The 11th Annual Bio-Ontologies Meeting - Bio

The 11th Annual Bio-Ontologies Meeting
Phillip Lord, Newcastle University
Nigam Shah, Stanford University
Susanna-Assunta Sansone, EMBL-EBI
Matthew Cockerill, BioMed Central
The Bio-Ontologies Meeting has existed as a Special Interest Group meeting at ISMB for more than a
decade, making it one of the longest running SIG meetings. The Bio-Ontologies SIG meeting has
provided a forum for discussion on the latest and most cutting edge research on ontologies. In this
decade, the use of ontologies has become mature, moving from niche to mainstream usage within
bioinformatics. As result, this year we are broadening the scope of SIG to include formal and informal
approaches to organising, presenting and disseminating knowledge in biology.
July 20, 2008
Colocated with ISMB 2008
Toronto, Canada
Keynote
This year's keynote speaker will be Professor Philip E. Bourne, from the Skaggs School of Pharmacy &
Pharmaceutical Sciences, UCSD.
Panel Session
This year’s panel session will have the following members:
•
•
•
•
Philip E. Bourne, UCSD
Helen Parkinson, European Bioinformatics Institute
Matt Cockerill, BioMed Central
Mark Wilkinson, University of British Columbia
Acknowledgements
We acknowledge the assistance of Steven Leard and all at ISCB for their excellent technical assistance.
We also wish to thank the programme committee for their excellent input and guidance – the
programme committee, organised alphabetically is:
Mike Bada, University of Colorado
Judith Blake, Jackson Laboratory
Frank Gibson, Newcastle University
Cliff Joslyn, Pacific National Laboratory
Wacek Kusnierczyk, Norwegian University of Science and Technology
Robin MacEntire, GSK
James Malone, EBI
Helen Parkinson, EBI
Daniel Rubin, Stanford University
Alan Ruttenberg, Science Commons
Susie M. Stephens, Eli Lilly
Robert Stevens, University of Manchester
Talks schedule
Start
8:50
9:00
End
9:00
9:20
Authors
9:20
9:40
9:40
10:00
C. Brewster et al
T. Tong et al
10:00
10:40
10:40
11:00
Coffee
T. Beck et al
11:00
11:20
J. Malone et al
11:20
11:40
C. Arighi et al
11:40
12:30
13:50
12:30
13:50
15:10
Phil Bourne
Lunch and Poster Session 1
Panel
15:10
15:30
A.X.Qu et al
15:30
15:50
R. Hoehndorf et al
15:50
16:10
W. Xuan et al
16:10
16:40
16:40
17:00
Coffee
J.B.L. Bard et al
17:00
17:20
R. Arp et al
17:20
18:30
Closing Remarks and Poster
Session 2
R. Kanagasabi et al
Title
Introduction
Ontology-Centric navigation of pathway
information mined from text
Issues in learning an Ontology from Text
GO-WORDS: An entropic Approach to
Semantic Decomposition of Gene Ontology
Terms
Using ontologies to annotate large-scale
mouse phenotype data
Developing an application focused
experimental factor ontology: embracing the
OBO Community
TGF-beta Signaling Proteins and the Protein
Ontology
Keynote talk
Philip E. Bourne, UCSD
Helen Parkinson, European Bioinformatics
Institute
Matt Cockerill, BioMed Central
Mark Wilkinson, University of British
Columbia
Tamoxifen to Systemic Lupus Erythematosus:
Constructing a Semantic Infrastructure to
Enable Mechanism-based Reasoning and
Inference from Drugs to Diseases
BOWiki: An ontology-based wiki for
annotation of data and integration of
knowledge in biology
PubOnto: Open Biomedical Ontology-Based
Medline Exploration
Minimal Information about Anatomy (MIAA):
a species-independent terminology for
anatomical mapping and retrieval.
Function, Role and Disposition in Basic
Formal Ontology
Posters
No.
1
Authors
2
Pascale Gaudet, Petra Fey, Eric Just, Sohel
Merchant, Siddhartha Basu, Warren Kibbe
and Rex Chisholm
Joanne Luciano, Lynn Schriml, Burke Squires
and Richard Scheuermann
Son Doan, Quoc Hung Ngo, Ai Kawazoe and
Nigel Collier
3
4
5
6
7
8
9
Joel Sachs, Cynthia Parr, Lushan Han, Taowei
Wang and Tim Finin
Philippe Rocca-Serra, Ryan Brinkman, Liju
Fan, James Malone and Susanna-Assunta
Sansone
Lynn Schriml, Katherine Phillippy, Aaron
Gussman, Anu Ganapathy, Sam Angiuoli,
Suvarna Nadendla, Victor Felix, Elodie
Ghedin, Owen White and Neil Hall
François Belleau, Peter Ansell, MarcAlexandre Nolin, Kingsley Idehen and Michel
Dumontier
Andrea Splendiani, Scott Piao, Yutaka Sasaki,
Sophia Ananiadou, John McNaught and anita
burgun
Mary Shimoyama, Melinda Dwinell and
Howard Jacob
Title
Case Studies of Simple Ontologies and Instance Data
for Biodiversity Informatics
Strain and phenotype annotation at dictyBase
The Influenza Infectious Disease Ontology (I-IDO)
Building and Using Geospatial Ontology in the
BioCaster Surveillance System
OBI: The Ontology for Biomedical Investigations
The Evolution of Controlled Vocabularies and
Community Development Ontologies in the Frame
Work of an Epidemiology Database
Bio2RDF's SPARQL Point and Search Service for Life
Science Linked Data
Compositional enrichment of bio-ontologies
Multiple Ontologies for Integrating Complex
Phenotype Datasets
10
Indra Neil Sarkar, Ryan Schenk, Holly Miller
and Catherine Norton
Identification of Relevant Biomedical Literature
Using Tag Clouds
11
Jason Greenbaum, Randi Vita, Laura M.
Zarebski, Alessandro Sette and Bjoern Peters
Ontology of Immune Epitopes (ONTIE)
Ontology-centric navigation of pathways mined from text.
Rajaraman Kanagasabai1, Hong-Sang Low2, Wee Tiong Ang1, Markus R. Wenk2, and
Christopher J. O. Baker1
1
Data Mining Dept. Institute for Infocomm Research, & 2Departmet of Biochemistry, Faculty of Medicine, NUS, Singapore,
ABSTRACT
Motivation: The scientific literature is the primary means for
the navigation of new knowledge as well as retrospective
analysis across merging disciplines. The state of the art in
this paradigm is a series of independent tools that have to be
combined into a workflow and results that are static representations lacking sophisticated query-answer navigation
tools. In this paper we report on the combination of text mining, ontology population and knowledge representation technologies in the construction of a knowledgebase on which
we deploy data mining algorithms and visual query functionality. Integrated together these technologies constitute an
interactive query paradigm for pathway discovery from fulltext scientific papers. The platform is designed for the navigation of annotations across biological systems and data
types. We illustrate its use in tacit knowledge discovery and
pathway annotation.
1
INTRODUCTION
A growing number of knowledge discovery systems incorporate text mining techniques and deliver insights derived
from the literature. Named entity recognition and relation
detection are primary steps. The products of such techniques
can take the form of automatically generated summaries,
target sentences or lists of binary relations between entities
[1] from abstracts for which subsequent networks can be
constructed [2] and visualized in graphs [3] in some cases
with predefined class directed-layouts [4]. The state of the
art in this paradigm is a series of independent tools that have
to be combined into a workflow and results that are static
representations lacking sophisticated query-answer navigation tools. To enhance the accessibility and search ability of
the insights derived from texts these instances of named
entities and relations should be associated with descriptive
metadata such as ontologies. Recent examples have shown
that instantiating ontologies with text segments can be
meaningful and useful in knowledge discovery projects [5].
In this paper we go a step further and mine instantiated ontology for transitive relations linking query terms and make
this available in the context of (i) tacit knowledge discovery
across biological systems; proteins, lipids and disease, and
*
[email protected]
(ii) in mining for pathway segments which can iteratively be
re-annotated with relations to other biological entities also
recognized in full text documents.
2
METHODS
The material for our analysis is full text scientific literature.
Details and efficiency of our text mining approach, the customization of the ontology to enable the pathway discovery
scenario and instantiation of the ontology are detailed below. We also outline pathway discovery algorithms used to
facilitate navigation of putative pathways and annotations.
2.1
Ontology Population
Ontology population was achieved through the coordination
of content acquisition, natural language processing and ontology instantiation strategies. We employed a content acquisition engine that takes user keywords and retrieves fulltext research papers by crawling Pubmed search results.
Retrieved collections of research papers were converted
from their original formats, to ascii text and made ready for
text mining by a customized document converter. Knowledgebase ‘instances’ are generated from full texts provided
by the content acquisition engine using the BioText toolkit.
http://datam.i2r.a-star.edu.sg/~kanagasa/BioText/
2.2
Knowledgebase Instantiation
Instantiation comprises of three stages: Concept Instance
Generation, for which we also provide a performance
evaluation, Property Instance Generation, and Population of
Instances. In the context of OWL-DL, Property Instances
are assertions on individuals which are derived from relations found in predicate argument structures in mined sentences.
2.2.1 Concept Instance Generation. Concept instances are
generated by first extracting the name entities from the texts
and then normalizing and grounding them to the ontology
concepts. Our entity recognizer uses a gazetteer that processes retrieved full-text documents and recognizes entities
by matching term dictionaries against the tokens of processed text, and tags the terms found [5]. The lipid name dictionary was generated from Lipid Data Warehouse that contains lipid names from LIPIDMAPS [6], LipidBank and
1
Kanagasa et al.
KEGG., IUPAC names, and optionally broad synonyms and
exact synonyms. The manually curated Protein name list
from Swiss-Prot (http://au.expasy.org/sprot/) was used for
the protein name dictionary. A disease name list was created from the Disease Ontology of Centre for Genetic Medicine (http://diseaseontology.sourceforge.net). Our normalization and grounding strategy is as follows. Protein names
were normalized to the canonical names entry in Swiss-Prot.
Grounding is done via the Swiss-Prot ID. For lipid names,
we define the LIPIDMAPS systematic name as the canonical name, and the LIPIDMAPS database ID is used for
grounding. Disease names are grounded via the ULMS ID.
To evaluate the performance of our named entity/concept recognition we constructed a gold standard corpus of 10 full-texts papers related to the Apoptosis pathway,
obtained from our content acquisition engine. We extracted
119 sentences and tagged the mentions of Protein name and
Disease name. In these sentences we annotated all valid
mentions of the two concepts and built the corpus. To
evaluate performance of named entity/concept recognition a
corpus without the concept annotations was passed to our
text mining engine and the concepts recognized. Our system
was evaluated in terms of precision and recall. Precision
was defined as the fraction of correct concepts recognized
over the total number of concepts output, and recall was
defined as the fraction of concepts recognized among all
correct concepts.
Table 1. Precision and recall of named entity recognition
Mentions
Named Entities
Disease
Lipid
Protein
Micro Average
Target
Returned
32
58
269
37
25
181
Precision
Recall
0.54
0.96
0.76
0.75
0.62
0.47
0.51
0.51
tences are instantiated as Datatype property instances. Several other Object property instances are also generated to
establish relations between, for example, LIPIDMAPS Systematic Name and its associated IUPAC Name, synonyms
and database ID. However, in this case the relation pairs are
generated directly from the Lipid Warehouse records requiring no text processing.
2.2.3 Population of Instances. In this step we collect all the
concept and property instances generated from the previous
two to instantiate the ontology. The concept instances are
instantiated to the respective ontology classes (as tagged by
the gazetteer), the Object Property instances to the respective Object Properties and the Datatype property instances
to the respective Datatype properties. We wrote a custom
script using the OWL programming framework, JENA API
http://jena.sourceforge.net/ for this purpose.
2.3
Ontology Extension
To facilitate the navigation of pathway information we
modified the existing lipid ontology [5] by incorporating
Protein concepts under two newly defined superconcepts
(i) Monomeric_Protein_or_Protein_Complex_Subunit and
(ii) Multimeric_Protein_Complex. This was achieved either
by importing protein entities found in Molecule Roles Ontology or by adding the names manually. In total, we incorporated about 48 protein class entities under these 2 concepts. Each protein entity relates to another via the property
“hasProtein_Protein_Interaction_with”. Each protein entity
then relates to a lipid entity via the property “interactsWith_Lipid”. These extensions facilitate query of protein-protein interactions derived from tuples found by the
text mining of full text documents.
Generic Pathway Algorithm
Let C
= source concept, C
source
The evaluation of entity recognition shown in Table 1
shows that our text mining achieved performance comparable to that of the state-of-the-art dictionary-based approaches. In our future work, we plan to make use of advanced entity recognition techniques, e.g. fuzzy term matching and coreference resolution, and also train our system on
larger corpora, to address these issues.
If C
= C
source
2
, output P as the pathway and stop.
target
Check if C
= D where D ∈ T = {D, P, R}, and D and R
source
are concept instances related via the object property P.
If C
= R where R ∈ T = {D, P, R}, add the edge C
target
be R, and go to Step 1.
Else, Let C
→ T* → C
source
to P.
target
source
= R where R ∈ T = {D, P, R}.
Check if C
source
If C
2.2.2 Property Instance Generation. Object property and
Datatype property instances are generated separately. From
the Lipid, Protein and Disease instances, four types of relation pairs namely Lipid-Protein, Lipid-Disease, ProteinProtein, Protein-Disease are extracted. For relation detection, we adopt a constraint-based association mining approach whereby two entities are said to be related if they cooccur in a sentence and satisfy a set of specified rules [5].
The relation pairs from the resulting sentences are used to
generate the Object property instances. The interaction sen-
= target concept, and P = null
target
= D where D ∈ T = {D, P ,R}, add the edge C
target
Else, Let C
be D, and go to Step 1.
source
← T* ← C
to P.
target
Fig. 1 Generic pathway algorithm for mining transitive relations.
2.4
Pathway Discovery Algorithm
We implemented a generic pathway discovery algorithm for
mining all object properties in the ontology to discover transitive relations between two entities. An outline of this algorithm is presented in Figure 1. Given two concept instances
Csource and Ctarget, the algorithm seeks to trace a pathway
between them using the following approach. First, the algorithm lists all object property instance triples in which the
Ontology-centric navigation of pathways mined from text.
domain matches Csource. Thereafter every listed instance is in
turn treated as the source concept instance and the related
object property instances explored. This process is repeated
recursively until Ctarget is reached or if no object property
instances are found. All resulting transitive paths are output
in the ascending order of path length. We implement a protein-protein interaction pathway discovery algorithm by
adding the following two simple constraints to the generic
algorithm: 1) the source and domain concepts are restricted
to be proteins, 2) only object property instances of hasProtein_Protein_Interaction_with are included.
3
Fig. 2 A tacit knowledge query in Knowlegator, searching for
links between two proteins in Apoptosis signaling.
PATHWAY MINING FROM LITERATURE
Knowlegtor [1] is a visual-query navigation platform for
OWL-DL ontologies which facilitates the construction of
concept level queries from OWL ontology constructs and
relays them to a reasoner to query a knowledgebase populated with A-box instances. We have integrated 2 new
pathway algorithms into the Knowlegator platform to facilitate literature driven tacit knowledge discovery and apply it
here as ‘pathway mining’.
In order to mine the instantiated ontology for the existence
of one or more pathways between user-specified proteins
the graphical features of Knowlegator permit users to drag
two protein onto the query canvas and then invoke a search
for transitive relations between these two concepts (Figure
2). Results from this search are returned as a list of possible
pathways each of which can rendered on the query canvas
as a chain of labeled concepts and instances illustrating the
linkage between the selected starting entities (Figure 3).
These pathways traverse multiple relation and data types,
namely, protein, lipid and disease names as well as provenance data i.e. individual sentences and document identifiers. Parent concept names are rendered along with instance
level names. By using a wide range of relations a deeper
search for tangible relations between entities is facilitated.
This is however beyond the scope of pathway analysis and
more in line with identifying evidence sources and illustrating causality or participants in a disease context. There exists however many such paths through the instantiated ontology and the user’s navigation experience may become
tedious, in particular when the user is confronted with significant sources of new material. Within the process of
knowledge discovery a more intuitive approach is the iterative overlay of new material on top of existing knowledge
that is queried in the first stage of the analysis. In this vain
we illustrate the overlay of lipid-protein interaction information on top of the protein-protein interaction information
displayed in the initial pathway discovery step. Knowlegtor
facilitates use of the second pathway algorithm for users
who wish to apply specific constrains on the pathway they
hope to find, be it based on protein-protein interactions,
protein-disease or protein-lipid interactions. Figure 4 shows
a pathway query for a protein-protein pathway with associated protein-lipid interactions, the results for which are
Fig. 3 Results of a query to the instantiated ontology using PI3K
kinase and P53 which both play roles in apoptosis signaling.
provided for the PI3-Kinase to BIM II fragment of the apoptosis pathway (Figure 5). In our initial trials we considered
the
following
3
known
protein pathways:
1. PI3K --> Akt(Akt1; Akt2; Akt3) --> Bad(Bad)
--> Bcl-xl(Bcl-xL:) --> Bid(Bid) --> Bax(Bax)
2. PI3K --> Akt(Akt1;Akt2;Akt3) --> Mdm2(Mdm2)
--> p53(p53) --> Puma(Puma) --> Bim(Bim)
3. PI3K --> Akt(Akt1;Akt2;Akt3) -->FoxO1(FoxO1)
--> FasL(FasL) --> FasR(FasR) --> Caspase-3
where PI3K is (PI3-kinase p85-alpha subunit; PI3-kinase
p85-beta subunit; PI3-kinase p110 subunit alpha; PI3kinase p110 subunit beta; PI3 kinase p110 subunit delta).
For each pathway, we ran our pathway discovery algorithm
with first protein as the source concept and the last protein
as the target concept. There were > 1000 paths linking the
soure and target proteins and the correct pathway was identified as the path that had the identical P2P interactions as in
the known pathway and in the same order, stipulated by our
domain expert. With our system we were able to identify 2
out of 3 known pathways exactly. The third pathway was
found except for the segment PN_FoxO1 => PN_FasL =>
PN_FasR. Instead our algorithm identified a shorter path
"PN_FoxO1 => PN_FasR indicating there system is able to
pick up nuances beyond common ‘textbook’ pathways.
3
Kanagasa et al.
and for verification of the value added by the system in their
discovery process, which in some contexts is subjective.
Pathway Fragment: PI3-Kinase to BIM II
PN_PI3-kinase_p110_subunit_alpha => PN_Akt1
PathWithLipid: PN_PI3-kinase_p110_subunit_alpha => Systematic_LN_ethanoic_acid =>
PN_Akt1
PathWithLipid: PN_PI3-kinase_p110_subunit_alpha => Systematic_LN__5Z_7E_par__3S_par_-9_10-seco-5_7_10_19_par_-cholestatrien-3-ol => PN_Akt1
Fig. 4 shows the query for an apoptosis pathway fragment involving PI3K kinase and P53 and for lipid-protein annotations.
The URL below links to three sample pathways found using
the transitive query - protein interaction algorithm.
http://datam.i2r.a-star.edu.sg/~kanagasa/pathway/index.html
4
DISCUSSION
PathWithLipid: PN_PI3-kinase => Systematic_LN_Paclitaxel => PN_Akt1
PN_Akt1 => PN_Mdm2
PathWithLipid: PN_Akt1 => Systematic_LN_ethanoic_acid => PN_Mdm2
PathWithLipid: PN_Akt1 => Systematic_LN_GalNAca1-3_Fuca1-2_par_Galb13GlcNAcb1-3Galb1-3GlcNAcb1-3Galb1-4Glcb-Cer => PN_Mdm2
PN_Mdm2 => PN_p53
PathWithLipid: PN_Mdm2 => Systematic_LN_1-Methoxy-3_4-didehydro-1_2_7__8_tetrahydro-psi_psi-caroten-2-one => PN_p53
PathWithLipid: PN_Mdm2 => Systematic_LN_ethanoic_acid => PN_p53
The challenge we address in our scenario is the aggregation
of tuples of normalized named entities from full text documents and the provision of these tuples as an interactive
query resource for pathway discovery. The overall workflow has a series of exchangeable components which make
it an attractive solution. In future we plan to evaluate the
benefits to the overall system of exchanging different components, such as the information extraction engine which
exports tuples to a tagged XML file. In the current system
ontology population takes an average of 1 min/document,
mining the protein-protein pathways takes on average of 45
secs, and the automated verification of a known pathway
took less than 1 sec. In addition to generating content and
modifying the ontology to support the instantiation of protein-protein interactions, we have deployed two data mining
algorithms within the Knowlegator platform. With Knowlegator’s drag and drop query paradigm users can generate
cross-discipline paths or stepwise extensions to existing
known pathways by adding annotations or alternate paths
e.g. that include lipids. Moreover the results can be returned
with concept labels as well as instance names to enhance the
semantics of knowledge discovery output. We envision that
users of our approach would have a specific set of pathways
in mind from a given biological domain and specify a body
of literature to be mined and from which relevant information would be instantiated to the ontology. Thereafter these
users would navigate outwards from known pathways selecting to augment them with information which is beyond
their domain expertise. Moreover the ontology captures sentence provenance so that as users can verify new information that they were not previously aware of. Whilst this is
preliminary work it shows that mining literature sources in
the context of existing knowledge domain can support scientists engaged in knowledge discovery around pathways.
As we move forward with this paradigm we acknowledge
that we become more dependent on domain experts for precise requirements, (pathways and corresponding corpora)
4
PathWithLipid: PN_Mdm2 => Systematic_LN_2-methyl-propanoic_acid => PN_p53
PathWithLipid: PN_Mdm2 => Systematic_LN_GalNAca1-3_Fuca1-2_par_Galb13GlcNAcb1-3Galb1-3GlcNAcb1-3Galb1-4Glcb-Cer => N_p53
PN_p53 => PN_Puma
PathWithLipid: PN_p53 => Systematic_LN_1-Methoxy-3_4-didehydro-1_2_7__8_tetrahydro-psi_psi-caroten-2-one => PN_Puma
PathWithLipid: PN_p53 => Systematic_LN_Phorbol => PN_Puma
PN_Puma => PN_Bim
PathWithLipid: PN_Puma => Systematic_LN_1-Methoxy-3_4-didehydro-1_2_7__8_tetrahydro-psi_psi-caroten-2-one => PN_Bim
PathWithLipid: PN_Puma => Systematic_LN_Phorbol => PN_Bim
Fig. 5 shows the results of a query for a pathway fragment involving PI3K kinase and P53 along with lipid-protein annotations.
ACKNOWLEDGEMENTS
A*STAR (Agency for Science and Technology Research,
Singapore). National University of Singapore's Office of
Life Science (R-183-000-607-712), the Academic Research
Fund (R-183-000-160-112) and the Biomedical Research
Council of A*STAR (R-183-000-134-305).
REFERENCES
[1] Jang H, etal. BioProber: software system for biomedical relation discovery from, Conf Proc IEEE Eng Med Biol Soc.
2006;1:5779-82
[2] Chen H. and Sharp BM, Content-rich biological network constructed by mining PubMed abstracts, BMC Bioinformatics
2004, 5:147-160.
[3] Garcia O, etal.., GOlorize: a Cytoscape plug-in for network
visualization with Gene Ontology-based layout and coloring,
Bioinformatics. 2007 Feb 1;23(3):394-6.
[4] Douglas SM, Montelione GT, Gerstein M, PubNet: a flexible
system for visualizing literature derived networks, Genome Biology 2005, 6:R80
[5] Baker C.J.O., etal. Towards ontology-driven navigation of the
lipid bibliosphere, BMC Bioinformatics, vol. 9(Suppl 1), 2008
[6] Sud M., etal. Subramaniam S: LMSD: LIPID MAPS structure
database. Nucleic Acids Res 2007, 35:D527-D532.
Issues in learning an ontology from text
Christopher Brewster*, Simon Jupp§, Joanne Luciano¶, David Shotton# Robert Stevens§§, and Ziqi Zhang
*University of Sheffield, §University of Manchester, ¶Harvard University, #University of Oxford
ABSTRACT
Ontology construction for any domain is a labour intensive
and complex process. Any methodology that can reduce the
cost and increase efficiency has the potential to make a major impact in the life sciences. This paper describes an experiment in ontology construction from text for the Animal
Behaviour domain. Our objective was to see how much
could be done in a simple and rapid manner using a corpus
of journal papers. We used a sequence of text processing
steps, and describe the different choices made to clean the
input, to derive a set of terms and to structure those terms in
a hierarchy. We were able in a very short space of time to
construct a 17000 term ontology with a high percentage of
suitable terms. We describe some of the challenges, especially that of focusing the ontology appropriately given a
starting point of a heterogeneous corpus.
1
INTRODUCTION
Ontology construction and maintenance are both labour
intensive tasks. They present major challenges for any user
community seeking to use sophisticated knowledge management tools. One traditional perspective is that once the
ontology is built the task is complete, so users of ontologies
should not baulk at the undertaking. The reality of ontology
development is significantly different. For some large,
widely used ontologies, such as the Gene Ontology
(Ashburner et al. 2000), a manual approach is effective even
if very expensive. For small, scientific communities with
limited resources such manual approaches are unrealistic.
This problem is all the more acute as research in many areas, including the life sciences, is moving to an e-science
industrialised paradigm.
The work presented in this paper concerns the semiautomatic construction of an ontology for the animal behaviour domain. The animal behaviour community has recognised the need for an ontology in order to annotate a number
of data sets. These data sets include texts, image and video
collections. In a series of workshops1, an initial effort has
been made to construct an ontology for the purposes of applying annotations to these data sets. The current Animal
Behaviour Ontology (ABO) has 339 classes and the top
level structure is shown in Figure 1.
*
1
While considerable effort has already gone into the construction of the Animal Behaviour Ontology, its limited size
raises the important question as to whether it is more appro-
To whom correspondence should be addressed.
For further details cf. http://ethodata.comm.nsdl.org/
Figure 1 Top level terms in the Animal
Behaviour Ontology
priate to slowly build an ontology entirely by hand, and
have its potential expansion led by user demand, or whether
to rapidly build a much larger ontology based on the application of a variety of text processing methods, and tidy or
clean the output. With community engagement comes
growth, but there is a question of stimulating engagement
through some critical mass of useful ontology. The former
approach is the standard approach and has been used
successfully in cases such as the Gene Ontology, but
becomes more challenging as the size and complexity of the
ontology increases. On the other hand, while much has been
written about automatic ontology learning, most such work
has been undertaken in non-biological domains, or in rather
abstract contexts (Cimiano et al. 2005; Brewster et al. 2007;
Navigli and Velardi 2004). Although such research is called
“ontology learning” in reality, given the limitations of Natural Language Processing, the outputs have been structured
1
Error! No text of specified style in document.
Language Processing, the outputs have been structured vocabularies organised in taxonomic hierarchies. This might
be considered a major defect if it were not that a) most ontologies are used for labelling/annotation purposes rather
than for computational inference, and b) a hierarchically
structured vocabulary based on the actual terminology used
by a community is a major step towards the creation of a
formal ontology. Thus in our view, the construction of formal ontologies of the type needed for driving semantic applications should be considered to involve a significant
manual step following the automated process (Luciano and
Stevens 2007; Stevens et al. 2007).
In the research reported here, we chose to see how far we
could go in the context of limited resources. We approached
the challenge as being one to construct a controlled or structured vocabulary as quickly as possible, with minimal effort,
and then allow subsequent efforts to clean up the output of
this exercise. At one level, we have tried to assess how
much effort is worth investing and what is the balance of
cost and benefit. A greater understanding of what is the best
and most effective methods will in the longer term not only
facilitate the creation of useful ontologies for scientific domains with limited resources, but will also facilitate the
growing issue of maintenance and upkeep of ontologies as a
whole.
2
2.1
METHODOLOGY
The Data Set
It has been argued elsewhere that the only effective way to
build representative ontologies for a given domain is
through the use of text corpora (Brewster, Ciravegna, and
Wilks 2001), and in our case we were able to have access to
a considerable corpus of journal articles from the journal
Animal Behaviour, published by Elsevier. This consisted of
articles from Vol 71 (2006) to Vol 74 (2007), containing
623 separate articles. We were given access to text, PDF
and XML versions together with a corresponding DTD. We
used the XML version for the procedures which are described below.
2.2
From text to ontology
1. Clean text was extracted from the XML files. Using the
information from the structured markup, we excluded all
author names, affiliations and addresses, acknowledgements, and all bibliographic information, except for the titles of the cited papers.
2. A number of stop word lists and gazetteers were used to
further remove noise from the data. We excluded person
names as noted above and also through the use of a gazet-
teer, animal names based on a short list derived from the
LDOCE2, and place names using another gazetteer.
3. A lemmatizer was used to increase coverage (Zhou,
Xiaodan Zhang, and Hu 2007). In some cases this generated
some noise due to imperfections in the lemmatizer but overall it reduced data sparsity.
4. Five different term extraction algorithms were applied as
described in (Ziqi Zhang et al. 2008). The chosen term recognition algorithms were ones that selected both single and
multi-word terms as we believe that desirable technical
terms are of both sorts. The algorithms were applied to each
subsection of the journal article as well as to the whole. This
allowed us to look at the terms from different sections of the
articles (abstract, introduction, materials and methods, conclusion, etc.). as we aimed to build an ontology of animal
behaviour, the terms found exclusively in the “Materials and
Methods” section were removed from further consideration.
Such terms are the subject of a different ontology.
5a. We then used a set of regular expressions to filter the
candidate terms. A regular expression was constructed that
looked for terms that ended in behaviour, display, construction, inspection, etc. It also included some very generic
regular expressions looking for terms that ended in -ing and
–ism. The regular expression used for term selection is
available on the website accompanying this research3.
5b. The step described in 5a. involved quite specific domain
knowledge. To have an alternative procedure that does not
involve any domain knowledge, we used a voting algorithm
to rank the terms and weight them for distribution across the
corpus. This was calculated by taking the mean rank for
each term and multiplying by the document frequency.
From the resulting rankings terms were selected for the subsequent steps (to parallel those extracted by the regular expression).
6a. There are a number of methods that can take a set of
terms and try to identify ontological (taxonomic) relations
between the terms (Cimiano, Pivk, Schmidt-Thieme, and
Staab 2005; Brewster 2007). Most methods suffer from low
recall. So in our approach we chose to use the method used
in the literature with highest recall – string inclusion. This
means that a term A B IS_A B, and A B C IS_A (B C and A
C) IS_A C. The resulting ontology was saved in the Web
Ontology Language (OWL).
6b. The same method as 6a. was applied to the output of 5b.
7a. and 7b. The resultant ontologies were then filtered for
their top level terms i.e. children of THING. A technique
used extensively in the ontology learning community is that
of using lexico-syntactic patterns (or Hearst patterns (Hearst
1992)) to either learn or test for a candidate ontological relation (Brewster et al. 2007). In this case, we tested each top
2
The Longman Dictionary of Contemporary English. Our thanks to Louise
Guthrie for providing this.
3
http://nlp.shef.ac.uk/abraxas/animalbehaviour.html
2
Error! No text of specified style in document.
level term in each ontology as to whether it was a kind of
behaviour, activity or action using the Internet as an external resource. Thus we constructed phrases such as the following: “behaviours such as biting” (found) or “behaviours
such as dimorphism” (not found).
3
RESULTS
Figure 2 Partial subtree from ontology at Step 6a.
which 2535 classes were top level. The ontologies mentioned here are available on the web site accompanying this
paper4. A screen shot of the sub tree concerning call from
ontology 6a. is shown in Figure 2.
The filtering process described in Step 7a. resulted in 383
top level terms being removed leaving 912 immediate descendants of OWL:THING. Top level classes that were filtered out by this method included terms such as stocking
referencing, holding, attraction, time, schooling, movement,
pacing, defending, smashing, loading, matricide. The parallel process in 7b. resulted in 649 top level classes being removed, leaving 1886.
A sample of the terms excluded by step 5a. has been
evaluated by a biologist (Shotton). Of the 56,000 terms excluded, a random sample of 3140 terms were manually inspected. Of these 7 verbs and 42 nouns were identified as
putative animal behaviour-related terms. These included
terms such as forage, strike, secretion, ejaculate, higher
frequency yodel, female purring sound, etc. The low number of significant excluded terms shows that our approach
has a Negative Predictive Value of 0.98, and a Recall of
0.905. We have yet to determine the precision of this approach due to the need for large scale human evaluation of
the selected terms.
4
A total of 64,000 terms were extracted from the whole corpus of 2.2 million words. From this the regular expression
extracted 10,335 terms. These included animal behaviour
terms, but also included non-animal behaviour terms. The
regular expression was designed to capture a large number
of terms such as begging, foraging, dancing, grooming, burrowing, mating. Due to its crudity it also picked up nonbehavioural terms with similar endings: -bunting, -herring,
dichromatism, dimorphism.
The ontology produced by Step 6a. resulted in an artefact
of 17776 classes, of which 1295 classes are top level (i.e.
direct children of OWL:THING). The ontology produced by
Step 6b. from the 10,335 terms selected by the voting algorithm in step 5b. resulted in an artefact of 13,058 classes, of
DISCUSSION
A key challenge in the process of learning an ontology from
texts is to identify the base units, i.e. the set of terms which
will be used as labels in the ontology’s class hierarchy. This
problem has been largely ignored in the NLP ontology
learning literature. The problem of constructing an ontology
from a data set such as the one we were using is that in effect there are a number of different domain ontologies represented in the text. In the case of our corpus from the journal Animal Behaviour, there existed terms reflecting experimental methods, animal names, other named entities
(places, organisations, people), etc in addition to behaviours. Such domains are obviously pertinent to animal behaviour (there are species specific behaviours), but the terms
exclusively from these domains belong to separate ontologies. The linking together of these separate domains within
one ontology is a further step in the process of ontology
building.
In order to construct an ontology of animal behaviour from
such a heterogeneous data set, one must focus the term selection as much as possible. In order to do this we used first
a manually constructed set of regular expressions, an approach which is dependant on domain expertise. As an alternative, for the sake of comparison, we selected the same
number of terms using the term recognition voting approach. The ontology generated by this latter approach re-
4
http://nlp.shef.ac.uk/abraxas/animalbehaviour.html
3
Error! No text of specified style in document.
sulted in less complexity because it included fewer multiword terms, which using our string inclusion method had
generated further intermediate concepts and a richer hierarchy when using the terms identified by regular expressions.
Our initial evaluation of the terms excluded by the regular
expressions shows that very few of the omitted terms were
significant from an expert’s perspective. Our approach will
tend to high recall and low precision so there are certainly a
significant number of terms included that would need subsequent manual exclusion. A brief consideration of Figure 2
shows a number of terms that would need to be excluded:
g_call, lower call, etc.
Nevertheless, the resulting ontologies, especially after filtering the top level terms, contains a large number of useful
taxonomic fragments even if there is quite a lot of noise.
Part of the principle of our approach, as noted in the Introduction, is that it is far easier to collect a large set of potentially significant ontological concepts automatically and
then eliminate the noise than to slowly build up a perfectly
formed but incomplete set of concepts but which inevitably
will exclude a lot of important domain concepts. Such an
artefact is far from a formal ontology but is nonetheless useful as a step towards a taxonomic hierarchy for the annotation of research objects, and as a stepping-stone to a more
formal ontology. While we still have to undertake a full
evaluation, initial assessments indicate the ontologies derived using the regular expressions are cleaner and of
greater utility.
The limitations of our approach may be summarised as
follows: a) there is a certain amount of noise in the resulting
ontologies (which we specify more precisely in future
work), b) some effort is involved in focussing the ontology
produced (i.e. to exclude terms that properly belong to another domain/ontology), c) the result is only taxonomic –
the use of string inclusion implies an ISA hierarchy although careful inspection shows that this is not always the
case.
The significance of our approach is that it is very quick
and easy to undertake. The results produced are very useful,
both in themselves as a knowledge discovery exercise in a
scientific domain, and as a stepping stone to a more rigorous
or formal ontology. The very low effort involved in the
process means that this type of data collection could be used
in all cases when building ontologies from scratch. We also
propose this approach as being a significant tool in ensuring
ontologies are up to date and are current with the terminology of a domain.
Future work will include applying the full Abraxas methodology (Brewster et al. 2007) to construct the richest possible
structure from the existing ontology. We plan a more extensive evaluation of the noise present i.e. terms that should be
excluded. At a more fundamental level, we need to consider
how appropriate it is to use terms derived from a corpus for
the building of an ontology in contrast to a formally and
rigorously hand built ontology.
ACKNOWLEDGEMENTS
We would like to thank Anita de Ward of Elsevier for making the text available from the Journal Animal Behaviour.
This work was supported by the AHRC and EPSRC funded
Archeotools project (Zhang), the Companions project
(www.companions-project.org) (IST-FP6-034434) (Brewster), and the Sealife project (IST-2006-027269) (Jupp).
REFERENCES
Ashburner, M. et al. 2000. Gene ontology: tool for the unification
of biology. The Gene Ontology Consortium. Nat Genet 25, no.
1:25-29.
Brewster, Christopher. 2007. Mind the Gap: Bridging from Text to
Ontological Knowledge. Department of Computer Science,
University of Sheffield.
Brewster, Christopher, Fabio Ciravegna, and Yorick Wilks. 2001.
Knowledge Acquisition for Knowledge Management: Position
Paper. In Proceeding of the IJCAI-2001 Workshop on Ontology
Learning,
Seattle,
WA
http://www.dcs.shef.ac.uk/~kiffer/papers/ontolearning.pdf.
Brewster, Christopher et al. 2007. Dynamic Iterative Ontology
Learning. In Recent Advances in Natural Language Processing
(RANLP 07), Borovets, Bulgaria.
Cimiano, Philipp, Aleksander Pivk, Lars Schmidt-Thieme, and
Steffen Staab. 2005. Learning Taxonomic Relations from Heterogeneous Sources of Evidence. In Ontology Learning from
Text: Methods, Evaluation and Applications, Frontiers in Artificial Intelligence,
IOS Press http://www.aifb.unikarlsruhe.de/WBS/pci/OLP_Book_Cimiano.pdf.
Hearst, Marti. 1992. Automatic Acquisition of Hyponyms from
Large Text Corpora. In Proceedings of the Fourteenth International Conference on Computational Linguistics (COLING 92),
Nantes, France, July 1992.
Luciano, Joanne S, and Robert D Stevens. 2007. e-Science and
biological pathway semantics. BMC Bioinformatics 8 Suppl
3:S3.
Navigli, Roberto, and Paula Velardi. 2004. Learning Domain Ontologies from Document Warehouses and Dedicated Websites.
Computational Linguistics 30, no. 2:151-179.
Stevens, Robert et al. 2007. Using OWL to model biological
knowledge. Int. J. Hum.-Comput. Stud. 65, no. 7:583-594.
Zhang, Ziqi, Jose Iria, Christopher Brewster, and Fabio Ciravegna.
2008. A Comparative Evaluation of Term Recognition Algorithms. In Proceedings of the Sixth International Conference on
Language Resources and Evaluation (LREC08), Marrakech,
Morocco.
Zhou, Xiaohua, Xiaodan Zhang, and Xiaohua Hu. 2007. Dragon
Toolkit: Incorporating Auto-learned Semantic Knowledge into
Large-Scale Text Retrieval and Mining. In Proceedings of the
19th IEEE International Conference on Tools with Artificial Intelligence (ICTAI) http://www.dragontoolkit.org/dragontoolkit.pdf.
4
GO-WORDS: An Entropic Approach to Semantic Decomposition of
Gene Ontology Terms
Tuanjie Tong, Yugyung Lee, and Deendayal Dinakarpandian*
School of Computing and Engineering, University of Missouri-Kansas City, Kansas City, Missouri 64110, USA
The underlying motivation of the work described in this
paper is to map annotations based on the Gene Ontology
(GO) (Ashburner, et al., 2000) to a semantic representation
that exposes the internal semantics of GO terms to computer
programs. The Gene Ontology (GO) views each gene product as being a structural component of a biological entity,
being involved in a biological process, and as having a molecular function. These three dimensions of component (C),
process (P) and function (F) are hierarchically refined into
several thousand subconcepts or GO terms for a finegrained description of gene products, and ultimately a representation of collective biological knowledge. The machinereadability of GO is based on explicit IS-A or PART-OF
relations between different GO terms (Fig. 1). The representation of each GO term in terms of a phrase in English is
primarily meant for human readability, and not machinereadability (Wroe, et al., 2003) (Fig. 1). For example, while
both humans and computer programs can understand that
‘Folic Acid Transporter Activity’ is one kind of ‘Vitamin
Transporter Activity,” only a human reader can appreciate
that proteins annotated with ‘Folic Acid Transporter Activity’ actually transport the vitamin folic acid. In other words,
the compositional semantics embedded within each GO
term is not currently accessible by computer programs; each
term per se is effectively a black box or meaningless string
of characters to computer programs.
It has been estimated that about two-thirds of GO terms
(Ogren, et al., 2004) contain another GO term as a substring
within it. For example, the GO term ‘Transporter Activity’
is a substring of several GO terms such as ‘Vitamin Transporter Activity’ and ‘Biotin Transporter Activity.’ In other
*
To whom correspondence should be addressed.
© Oxford University Press 2003
Human View
Vitamin Transporter Activity
Folic Acid Transporter Activity
Machine View
Vitamin Transporter Activity
IS-A
1 INTRODUCTION
words, many GO terms are combinations of distinct semantic units, as opposed to being a completely new concept.
The compositional nature of GO terms has the side effect of
resulting in a combinatorial increase in the size of GO. For
example, ‘Folic Acid’ appears in 12 different GO terms like
‘Folic Acid Transport,’ ‘Folic Acid Binding,’ and ‘Folic
Acid Transporter Activity.’ Similarly, the vitamin Biotin
appears in 23 GO terms, including 6 terms identical to that
IS-A
ABSTRACT
The Gene Ontology (GO) has a large and growing number of
terms that constitute its vocabulary. An entropy-based approach is presented to automate the characterization of the
compositional semantics of GO terms. The motivation is to
extend the machine-readability of GO and to offer insights
for the continued maintenance and growth of GO. A prototype implementation illustrates the benefits of the approach.
Folic Acid Transporter Activity
Fig. 1. The internal semantics of GO terms are visible
to humans but not to computer programs
for Folic Acid except for the replacement of ‘Folic Acid’
with ‘Biotin,’ e.g., ‘Biotin Transport,’ ‘Biotin Binding’ and
‘Biotin Transporter Activity.’ This phenomenon has been
one of the motivating factors behind the GO Annotation
Tool (GOAT) (Bada, et al., 2004) and the Gene Ontology
Next Generation (GONG) project (Wroe, et al., 2003),
which suggested having multiple intersecting hierarchies,
with a proposed evolution towards a DAML+OIL representation. Reasons for studying the compositional nature of GO
are to suggest missing relations (Mungall, 2004; Ogren, et
al., 2004), suggest new terms (Lee, et al., 2006; Ogren, et
al., 2004), increase computability of GO (Doms, et al.,
2005; Ogren, et al., 2004; Wroe, et al., 2003), and for providing models for GO-based analysis of natural language
processing of text (Blaschke, et al., 2005; Couto, et al.,
2005; Doms and Schroeder, 2005).
One way to discretize GO is to represent it as a language
consisting of progressive concatenation of tokens in the
form of regular expressions. An example of this is Obol
(Mungall, 2004), a language that exploits the regularity of
GO term names to represent it in Backus-Naur format.
However, this is applicable to only a subset of all GO terms.
In this paper, we use an entropic approach for the analysis
of regularity of GO term nomenclature. We show how this
1
T. Tong et al.
may be used to detect sets of GO terms sharing similar semantics. The decomposition of GO terms presented here
also suggests a way to minimize the complexity of GO.
2 METHODS
The general principle is to find clusters of GO terms sharing
similar semantic structure. Entropy (see below) is used to
find GO terms that share consistent location of a specific
token (word) within them. Each cluster is evaluated and a
corresponding semantic rule created.
Analysis of position-dependent conservation of GO tokens
Each GO term (version Feb 16th, 2006), including synonyms, was tokenized on white space into a sequence of individual words. For example, the GO term “L-amino acid
transport” is tokenized as “L-amino” + “acid” + “transport.”
Entropy (Shannon 1950) is used to measure the regularity in
location of each token within all GO terms:
l
EP t =
∑-
i = 1
p it log p it
where EPt is the positional entropy of token t, l is the length
(in number of tokens) of the longest GO term or synonym
t
that token t is observed to occur in, and pi is the probability
of finding token t at position i. If the logarithm is in base 2,
then entropy can be quantified in terms of bits. Recognizing
that gene product and molecule names embedded in GO
terms consist of a variable number of tokens, we choose to
note the position of each token relative to both the beginning and end of each GO term. For example, the token “acceptor” almost always occurs at the end of a GO term (with
the sole exception of the term “electron acceptor activity”).
Thus, it is uniformly the first term when counted from the
end of a GO term, with a resulting low positional entropy of
0.08 with respect to the end (EPE). In contrast, this token
has a highly variable position when counted from the start
of a GO term (as many as 15 different locations) resulting in
a high positional entropy (EPS) value of 3.3. If we focus
only on an EPS value, we would miss its positional conservation, i.e., tendency to occur at the very end of GO terms.
Since Shannon entropy is based only on proportions, it does
not distinguish between token distributions like [1, 1] (token
found once at the first position, and once at the second) and
[100, 100] (token found a hundred times each at the first and
second positions). Both would yield an entropy value of 1
bit even though there are only 2 occurrences of the former
and 200 of the latter. To distinguish between such tokens,
the absolute numbers of occurrence at a given distance from
either the start or end of GO terms are also recorded. The
calculated entropies are then ‘normalized’ (NEP) by adding
0.1 to the calculated value and dividing by the total number
of occurrences. Division of the entropic value by the total
number of occurrences yields lower values for a higher
2
number of token occurrences. The addition of 0.1 bit helps
to distinguish between tokens having an entropy of zero but
differing in their frequency of occurrence within GO terms.
For the above examples, this would yield values of (0.1/2 =
0.05) and (0.1/200 = 0.0005) respectively, thus yielding a
lower NEP value (implying higher degree of positional conservation after correction for more frequent occurrence) for
the more frequent token.
Semantic mapping rule generation
Tokens with low positional entropy, high number of occurrences or low normalized positional entropy are used as a
starting point for the generation of rules. For each such token, the corresponding set of GO terms is verified for semantic uniformity and a corresponding rule generated. This
takes minimal time as the majority of terms in a set follow
the same pattern. For example, ‘binding’ is a token that has
much lower entropy when measured from the end (0.28 bit)
than from the beginning (2.16 bits). The vast majority, 1544
out of 1597, of GO terms containing the token ‘binding’ end
General Steps
Mi
GO termj
Example
GO Annotation
Folate Transporter 1
Folic Acid Transporter Activity
Semantic
Tokenization
Mi
Ej + Rk
Folate Transporter 1
Folic Acid + Transporter Activity
Mapping
Rule
Mi
Rk
Ej
Folate Transporter 1
Transports
Folic Acid
MachineProse Assertion
Fig. 2. Mapping a GO annotation to a discretized triplet.
The general procedure is shown on the left together with
a specific example on the right
with it. 1524 of these are of the general form ‘Entity’ +
‘binding’ where ‘Entity’ represents one or more tokens in
succession representing a single concept. The Entity most
often specifies a molecule, and sometimes a structural component. The 20 exceptions include terms like ‘Protein domain specific binding’ and ‘regulation of binding.’ Thus, the
discretizing rule applicable to gene products {Mi} annotated
with these GO terms may be stated as ‘Mi binds Entity.’ In
other words, each corresponding GO term (e.g. Zinc Binding) is decomposed into a relational term (e.g. Binds) and
the embedded concept (e.g. Zinc). Thus, if the protein “40S
ribosomal protein S27” is annotated with the GO term ‘Zinc
Binding,’ then the corresponding discretized semantic form
is ‘40S ribosomal protein S27 Binds Zinc.’ Fig. 2 summarizes the general procedure with another example. Triplets
of this form correspond to MachineProse assertions
(Dinakarpandian, et al., 2006) and can contribute to an incremental knowledge-base distinct from paper publications.
3 RESULTS & DISCUSSION
GO-WORDS: An Entropic Approach to Semantic Decomposition of Gene Ontology Terms
GO-WORDS browser
Fig.3. Browser for analyzing tokens/words found within GO
terms. Columns 2 and 5 are measures of positional variation
of each token within GO terms, column 1 indicates whether
position in each row is with respect to the beginning or end
of corresponding GO terms, column 3 shows name of token,
and column 4 shows number of GO terms it is found in.
Tokenizing GO resulted in a 9152 unique tokens from a
total of 37,403 terms (20115 canonical + 17288 synonym
terms). Each token occurred 13.7 times on average. The
most frequent token was found to be “activity,” occurring a
total of 8891 times. In contrast, almost half the tokens
(4204), e.g. “xylem,” occurred only once. We implemented
a browser (Fig. 3) to analyze position-wise frequencies and
entropy of GO-terms. EP stands for entropy. The suffix S,
as in EPS indicates that positions were counted from the
beginning of the string, whereas the suffix E, as in EPE,
indicates that positions were counted backwards from the
end of the string. The prefix N indicates normalization (see
Methods above). Each token was analyzed using multiple
metrics. For example, Table I shows that the token ‘negative’ has the lowest positional entropy because it occurs
most of the time at the beginning of a GO term (1351 out
1358 occurrences with a corresponding EPS=0.055, and
normalized EPS=0.00004). In contrast, the token ‘oxidoreductase’ (not shown) has the highest positional entropy
(EPE=3.854;NEPE=0.019) because its 212 occurrences are
spread over 29 different positions within GO terms like
‘oxidoreductase activity, acting on paired donors, with incorporation or reduction of molecular oxygen, reduced
pteridine as one donor, and incorporation of one atom of
oxygen.’ Clearly, it is potentially easier to map GO terms
containing the token ‘negative’ than ‘oxidoreductase’ to a
machine-readable representation.
The GO-WORDS browser is a useful tool to gain insights
into the composition of GO terms. With respect to this paper, we focused on using it mainly to create semantic mapping rules. Thus, tokens with low values of NEPE (observed
range=0.00004 – 0.562) (Table I) and a large number of
occurrences were used to select GO terms for semantic
mapping to an assertion representation.
Given a token and position either from the beginning or end
of a string, the GO-WORDS browser lists all GO terms and
synonyms that share the token at a given position. For example, the token ‘transporter’ occurs second from the end
(517 out of 650) in GO terms like the following:
name: L-ornithine transporter activity
name: S-adenosylmethionine transporter activity
exact_synonym: S-adenosyl methionine transporter activity
name: adenine nucleotide transporter activity
name: spermine transporter activity
name: sulfite transporter activity
Table I. Tokens with lowest normalized positional entropy
Token
Normalized
Token
Entropy activity
Normalized
Entropy nepe=0.000
dehydrogenase
neps=0.001
negative
neps=0.000
cell
nepe=0.001
positive
neps=0.000
complex
nepe=0.001
metabolism
nepe=0.000
metabolism
neps=0.001
activity
neps=0.000
receptor
neps=0.001
binding
nepe=0.000
biosynthesis
neps=0.001
regulation
neps=0.000
transporter
nepe=0.001
of
neps=0.000
binding
neps=0.001
of
nepe=0.000
formation
neps=0.002
nepe=0.002
biosynthesis
nepe=0.000
ligand
pathway
nepe=0.000
catabolism
neps=0.002
regulation
nepe=0.001
transport
nepe=0.002
formation
nepe=0.001
cell
neps=0.002
anabolism
nepe=0.001
synthesis
neps=0.002
synthesis
nepe=0.001
acid
nepe=0.002
differentiation
nepe=0.001
proliferation
nepe=0.002
catabolism
nepe=0.001
acceptor
nepe=0.002
receptor
nepe=0.001
degradation
neps=0.002
breakdown
nepe=0.001
exocytosis
nepe=0.002
degradation
nepe=0.001
anabolism
neps=0.002
The general pattern for the above examples is “Entity transporter activity.” Thus, the mapping rule applicable to gene
products {Mi} annotated with these GO terms may be stated
as ‘Mi transports Entity,’ where entity is presumed to be the
prefix of ‘transporter activity.’ This assumption is true in
420 of the 440 cases. Exceptions to the rule include terms
like “siderophore-iron (ferrioxamine) uptake transporter
activity” and “transporter activity.” In the former, only a
subset of the prefix of “transporter activity” represents an
3
T. Tong et al.
Entity, i.e, the word ‘uptake’ doesn’t conform to the same
pattern. The latter is the parent term representing the abstract concept of ‘transporter activity.’
The GO token entropic measure helps in clustering terms
that share a token at the same relative position. Based on the
general patterns ‘Entity binding’ and ‘Entity transporter
activity,’ 23780 and 903 annotations respectively were
mapped to discretized triplets. However, the entropic analysis is based on the naïve assumption that each token represents a concept. In reality, names of entities often consist of
a variable number of words strung together, e.g., lipoprotein
lipase. Measuring the positional entropy of a token from
either end helps mitigate this problem to an extent, but only
to an extent. In particular, GO terms where the token of interest is flanked by entities of variable length will not show
a peak in the positional distribution. Further, since it is
based purely on a textual approach (no prior semantics),
manual verification is required to find sub-concepts that are
made up of contiguous tokens.
4 CONCLUSION
This paper has presented and addressed the advantages of a
discretized triplet representation of GO annotations and a
partially automated approach for doing so. In future, we
intend to extend the approach to the entire Gene Ontology,
combine information from other sources, and devise a sophisticated search interface that shall incorporate the MachineProse relation ontology (Dinakarpandian, et al., 2006).
The number of terms in GO has been rapidly growing since
its inception (Ashburner, et al., 2000). The total number of
terms has grown from 4507 in 2000 to more than 20,000 in
Feb 2006 (Gene Ontology Consortium). One reason is a
richer description, but redundancy of nomenclature is also a
factor. As GO is continuously revised (terms becoming obsolete, renamed and rearranged), maintaining its semantic
integrity is quite challenging. This paper suggests an approach to a leaner GO that is both people and machine
friendlier by allowing annotations to be built from reuse of
semantically defined building blocks. This would lessen the
growth rate of GO, with the resultant smaller size helping in
ensuring uniformity and semantic consistency of GO. The
benefits would be easier maintenance of GO and higher
semantic transparency. In the interim, a triplet view of GO
annotations offers a pragmatic solution. A potential advantage is to facilitate searches specified as a set of triplets,
occupying the middle ground between a natural language
interface and a keyword-based one.
Since a large number of entities within GO are general or
specific names of molecules, extracting the embedded molecular ontology would be a useful adjunct. Using other
ontologies like ChEBI (ChEBI) and completed mappings
between GO and other ontologies (Johnson, et al., 2006)
would be useful in this regard.
4
ACKNOWLEDGEMENTS
We would like to acknowledge funding support for this
work from the University of Missouri-Kansas City Research
Board to DD (FRG 2006) and from the University of Missouri Research Board to YL (UMRB 2005).
REFERENCES
Ashburner, M., Ball, C.A., Blake, J.A., Botstein, D., Butler, H.,
Cherry, J.M., Davis, A.P., Dolinski, K., Dwight, S.S., Eppig, J.T.,
Harris, M.A., Hill, D.P., Issel-Tarver, L., Kasarskis, A., Lewis, S.,
Matese, J.C., Richardson, J.E., Ringwald, M., Rubin, G.M. and
Sherlock, G. (2000) Gene ontology: tool for the unification of biology. The Gene Ontology Consortium, Nat Genet, 25, 25-29.
Bada, M., Turi, D., McEntire, R. and Stevens, R. (2004). Using
Reasoning to Guide Annotation with Gene Ontology Terms in
GOAT. SIGMOD Record, 33(2).
Blaschke, C., Leon, E.A., Krallinger, M. and Valencia, A. (2005)
Evaluation of BioCreAtIvE assessment of task 2, BMC bioinformatics, 6 Suppl 1, S16
ChEBI
Chemical
Entities
of
Biological
Interest.
http://www.ebi.ac.uk/chebi/
Couto, F.M., Silva, M.J. and Coutinho, P.M. (2005) Finding genomic ontology terms in text using evidence content, BMC bioinformatics, 6 Suppl 1, S21.
Dinakarpandian, D., Lee, Y., Vishwanath, K. and Lingambhotla,
R. (2006) MachineProse: an Ontological Framework for Scientific
Assertions, J Am Med Inform Assoc, 13, 220-232.
Doms, A., Fuche, T., Burger, A. and Schroeder, M. (2005) How to
Query the GeneOntology. Proceedings of Symposium on Knowledge Representation in Bioinformatics. Finland.
Doms, A. and Schroeder, M. (2005) GoPubMed: exploring PubMed with the Gene Ontology Nucleic Acids Res., W783-W786.
Gene Ontology Consortium.
Gene Ontology Annotations.
http://www.geneontology.org/
Johnson, H.L., Cohen, K.B., Baumgartner, J.W.A., Lu, Z., Bada,
M., Kester, T., Kim, H. and Hunter, L. (2006) Evaluation of Lexical Methods for Detecting Relationships Between Concepts from
Multiple Ontologies. Pacific Symposium on Biocomputing. 28-39.
Lee, J.B., Kim, J.J. and Park, J.C. (2006) Automatic extension of
Gene Ontology with flexible identification of candidate terms,
Bioinformatics, 22, 665-67
Mungall, C.J. (2004) Obol: integrating language and meaning in
bio-ontologies, Comparative and Functional Genomics, 509-520.
Ogren, P.V., Cohen, K.B., Acquaah-Mensah, G.K., Eberlein, J. and
Hunter, L. (2004) The compositional structure of Gene Ontology
terms, Pacific Symposium on Biocomputing, 214-225.
Shannon, C.E. (1950) Prediction and entropy of printed English.
The Bell System Technical Journal, 30, 50-64
Wroe, C.J., Stevens, R., Goble, C.A. and Ashburner, M. (2003) A
methodology to migrate the gene ontology to a description logic
environment using DAML+OIL, Pacific Symposium on Biocomputing, 624-635
Using ontologies to annotate large-scale mouse phenotype data
Tim Beck*, Ann-Marie Mallon, Hugh Morgan, Andrew Blake and John M. Hancock
MRC Harwell, Harwell, Oxfordshire, OX11 0RD, U.K.
ABSTRACT
The annotation of mouse phenotype data generated during a
large-scale primary phenotyping project is underway. Utilising OBO ontologies, a framework has been developed
which incorporates two existing annotation approaches to
form coherent and precisely defined descriptions of phenotyping procedures, the parameters of procedures and the
data derived for each parameter. We introduce the storage
of combinatorial phenotype ontology annotations at the database level with the use of interface controlled vocabularies, incorporating compound phenotype ontology terms, to
assist with phenotype capture at the point of data entry and
subsequent database querying.
1
INTRODUCTION
Two differing approaches can be adopted to annotate phenotype data with bio-ontologies. Either a single dedicated
ontology of compound terms can be employed or an annotation can be built using terms from a number of distinct ontologies to form a more complex expression to describe an
aspect of an organism’s phenotype [1]. The Mammalian
Phenotype (MP) ontology [2] is an example of a single dedicated phenotype ontology and the PATO model [3] of defining phenotypes in terms of an entity (E) which has a quality (Q) to build E+Q annotations is an example of the combinatorial approach. PATO is an ontology of species neutral phenotypic qualities and as such lends itself to the formation of comparable cross-species and cross-database phenotypic statements. Using the mouse kinked tail dysmorphological phenotype as an example, MP defines this phenotype
using the single term kinked tail (MP:0000585) and PATO
is used to assign a quality to the mouse anatomical entity
defined by the Mouse Anatomy (MA) ontology to form the
annotation E: tail (MA:0000008) and Q: kinked
(PATO:0001798).
MP has been widely implemented within database resources
with the Mouse Genome Informatics and the Rat Genome
Database providing associations between genes and MP
terms. However, although recently used for the description
of phenotypes observed during zebrafish screens [4], there
has, up to now, not been any such comprehensive imple*
To whom correspondence should be addressed.
mentation of the PATO combinatorial approach within
mammalian phenotype related informatics resources.
The European Mouse Disease Clinic (EUMODIC,
http://www.eumodic.org) is a major European project which
is undertaking a primary phenotype assessment of up to 650
mouse mutant lines derived from embryonic stem (ES) cells
developed in the European Mouse Mutagenesis
(EUCOMM) project. The phenotype assessment consists of
a selection of Standard Operating Procedures (SOPs) from
the European Mouse Phenotyping Resource of Standardised
Screens (EMPReSS, http://empress.har.mrc.ac.uk) [5] organised into two primary phenotyping pipelines. There is a
wide range of screens collecting phenotype data from the
mouse biological domains of morphology and metabolism;
cardiovascular; bone; neurobehavioral and sensory; haematology and clinical chemistry and allergy and immunity. As
a result of carrying out an individual SOP either quantitative
data (e.g. blood pressure measurement), qualitative data
(e.g. coat color description) or a combination of quantitative
and qualitative data (e.g. cornea opacity description and the
precise opacity level measurement) can be returned. The
date derived from carrying out the phenotyping pipelines are
stored in the EuroPhenome mouse phenotyping resource
(http://www.europhenome.eu) [6].
In order to unify the reporting of results from unrelated
mouse experimental procedures from different EUMODIC
research institutions, the data is converted to XML format
from the local Laboratory Information Management System
(LIMS) before submission to EuroPhenome. Each XML
file complies with the Phenotype Data XML (PDML) schema. PDML builds upon the Minimal Information for Mouse
Phenotyping Procedures (MIMPP), a minimum information
checklist which is under development to cover all mouse
phenotyping domains. MIMPP is a member of the Minimal
Information for Biological and Biomedical Investigations
(MIBBI) project whose goal it is to ensure descriptions of
methods, data and analyses support the unambiguous interpretation and reuse of data [7].
Each SOP has a number of parameters which define the type
of data to be recorded for a specific component of the SOP.
Either the parameter value will contain quantitative data, for
which the SI unit is specified, or it will be qualitative data.
1
T.Beck et al.
Qualitative data has traditionally been recorded using freetext or from a limited set of options. The application of
phenotype ontologies in the capture of qualitative data will
lead to more coherent and descriptive datasets. The structure of the PATO quality hierarchy lends itself favorably to
the annotation of parameters and the subsequently derived
data. For example, a parameter could be coat hair texture
and as a result of carrying out the SOP it is found to be
greasy. The quality greasy [is a] texture quality within
PATO, as are other potential mouse coat hair textures.
2
METHODS
The ontological annotation of mammalian phenotype data
was undertaken on three levels: the annotation of SOPs; the
annotation of individual SOP parameters and the annotation
of the data derived for each parameter. A distinction is
drawn between qualitative and quantitative phenotype data
as the annotation of these two classes of data is handled
differently.
2.1
SOP and parameter annotation
The SOPs were annotated using high-level MP terms to give
a general description of the procedure and provide a global
summary of all parameters within the SOP. The individual
parameters for each SOP were defined using the E+Q combinatorial approach in collaboration with scientists with
expert knowledge in each domain. European institutions
participating in EUMODIC record their primary phenotype
data using their in-house Laboratory Information Management System (LIMS), however the list of parameters for
each SOP is standardised. An overriding factor in the
process of parameter definition, especially while defining
parameters for qualitative data, was making the parameters
intuitive to the phenotyping scientists who would be interacting with the local LIMS. A desirable situation with respect to data accuracy and consistency would be one where
original LIMS entries could be imported directly into the
EuroPhenome data schema, therefore requiring that noninformaticians should be able to work seamlessly with ontologies. Given the large number of entries into the local
LIMS which would be required for a single SOP during the
lifetime of EUMODIC and the associated time cost, it was
essential that the practical implementation of ontology terms
to define parameters was accessible to phenotyping scientists. As a result of this process it was discovered that ontology classes and metadata did not exist to define anatomical entities using terminology that was understandable to
phenotyping scientists. These omissions were dealt with by
either proposing new terms for the MA ontology, submission of synonyms of existing terms or requests for term definitions.
2
2.2
Data annotation
Qualitative data, for example dysmorphology data, requires
the objective analysis of data at the point of data entry. Qualitative phenotypes, for example variations in coat colours,
are compared to wild-type mice and the researcher responsible for making the comparison must first make the decision as to whether a mouse is different and if it is, how it is
different. The use of ontologies in capturing qualitative data
at the point of data entry is desirable, since it would reduce
the ambiguity associated with interpreting free-text and the
subsequent mapping to an ontological structure. For this
reason the allowed values that could be assigned to a qualitative parameter were restricted to PATO qualities, specifically qualities that were child terms to the parameter defining quality. This process, in unison with the definition of
parameters, was carried out in collaboration with phenotyping scientists.
2.3
Interface parameter annotation using compound terms
The coherent and precisely defined E+Q structure of parameters and values lead to a marked increase in the number
of parameters to be evaluated for each qualitative SOP.
Additionally the decomposition of some community standard phenotypic terms to E+Q phenotype statements proved
problematic for phenotyping scientists to relate to during
data entry (see belly spot example below). For these reasons interface parameters were defined for qualitative SOPs
whereby intuitive compound MP terms defined parameters
and also could be assigned as values for parameters, instead
of simply assigning a child PATO quality to a parameter
PATO quality. The interface lists, while ensuring all value
options were restricted and so eliminating the need for freetext, also ensured that all interface parameters values were
mappable to the original formal E+Q parameters.
3
RESULTS
The need to develop interface controlled vocabularies that
contain compound phenotype terms which are intuitive for
use by mouse phenotyping scientists at the point of qualitative data entry presents an important data capture question.
If the phenotype data is to be accessible how should it be
contained within the database and subsequently queried and
presented to users? All EMPReSS SOP parameters are currently in the process of being defined using E+Q terms. In
addition to anatomical entities the ontological domains of
biological chemicals (CHEBI) and behaviors are also associated with qualities. Where qualitative data is concerned
the E+Q annotation is stored directly in the database.
Discussions with scientists during this practical ontology
annotation process has revealed that there is a preference for
Using ontologies to annotate large-scale mouse phenotype data
interacting with the database, either at the points of data
entry or querying, via community standard compound phenotype ontology terms where complex qualitative phenotypes are concerned. It is recognised that for some compound terms, when deconstructed into E+Q format, they
may lose their biological meaning. For example the term
belly spot (MP:0000373) is deconstructed to spotted
(PATO:0000333) [has quality] white (PATO:0000323) [inheres in] coat hair (MA:0000155) [part of] abdomen
(MA:0000029) (G.V. Gkoutos, personal communication).
A solution, as will be implemented within EuroPhenome, is
to store the phenotype in the database in the deconstructed
format but allow entry of the data and subsequent querying
via the compound term, so in this example belly spot. The
process of term mapping allows the interface parameter lists
to be implemented within local LIMS and then the phenotype to be imported into the EuroPhenome database in a
format which complies with the E+Q parameter lists.
As the use of bio-ontologies to define mouse phenotype
observations becomes increasingly commonplace it is essential that the ontologies are accessible and understandable by
those scientists who will make use of them and benefit from
their implementation the most. This demographic is no
longer restricted to ontologists or bioinformaticians, who
will continue to play an essential role in developing and
maintaining ontologies, but includes the “wet-science” researchers who will want to query large data sets using meaningful ontological terms and relationships in order to find
phenotypes of interest. A specific example taken from the
EUMODIC project would be scientists from secondary phenotyping clinics who will want to identify individual mice
exhibiting relevant mutant phenotypes from EuroPhenome
which will then undergo secondary phenotyping procedures.
These researchers will also become increasingly responsible
for entering their data into databases, albeit with appropriate
quality control mechanisms in place, so the descriptive
power of ontologies must be exploited to ensure they are as
scientist friendly as possible. We have identified a number
of omissions of terms from MA, for example nose skin,
which were regarded as essential for the precise categorisation of phenotypes. In other cases existing terms were not
intuitive to scientists and synonyms were suggested, for
example hind paw as a synonym of foot (MA:0000044) and
skull as a synonym of head bone (MA:0000576). Terms
were also identified which required definitions in order to
convey any useful meaning, for example foot digit 1
(MA:0000465) and hand digit 4 (MA:0000457). Our suggestions were passed onto MA curators. It is only through
the practical application of phenotype ontologies that omissions and potential improvements such as these will be identified.
4
DISCUSSION
We have described the ongoing efforts within the EuroPhenome mouse phenotyping resource to implement both the
MP and the E+Q combinatorial approach to systematically
annotate real mouse phenotypes, derived from community
approved SOPs, on a large scale. The three levels of annotations sees the marrying together of the two different phenotype annotation approaches into a framework that facilitates both data accessibility to mouse scientists using familiar terminology and also cross-database and cross-species
phenotype statement comparisons through the storage of
phenotypes in the E+Q format at the database level. Future
interfaces for the querying of EuroPhenome data will exploit mappings between MP and E+Q terms to accommodate the direct retrieval of E+Q annotations in addition to
querying via MP.
Currently only qualitative phenotypes are annotated with
ontologies. Quantitative data for baseline and mutant strains
across all phenotyping centers are entered into EuroPhenome. Comparative analyses between these two sets of data
allude to statistically significantly differences. Those mutant mice, which display significantly different values, are
then objectively annotated with MP and E+Q terms. The
annotation of quantitative data is therefore dynamic depending on the statistically significant characteristics queried. It
therefore follows that as sample sizes increase with the
amount of data in the EuroPhenome database (as the result
of future mouse strains undergoing phenotyping), the confidence in statistically deduced phenotype ontology annotations will in turn increase.
The work described here involving ontologies has focused
on the use of high-level definitions of SOPs, the definition
of parameters recorded as a result of carrying out the SOP
and the description of the data derived for each parameter.
Current research is focused on the development of an assay
ontology which will provide coherent definitions of each
individual procedural component contained within a SOP.
The context of a specific phenotype E+Q annotation would
be defined with the inclusion of this procedural data into a
phenotype data capture schema as illustrated in Fig 1.
Where qualitative data is available this would be stored
within the database and qualifies the phenotype quality value, which is a PATO child term of the parameter quality.
3
T.Beck et al.
7.
Parameter
E:tail (MA) + Q:length (PATO)
[assayed by]
SOP
measurement of tail length by ruler [unit] cm (UO)
[returns value]
Value
Q:increased length (PATO)
[qualified by]
Qualitative data
15
Fig 1. Phenotype and procedural data capture schema to
describe an instance of tail length. An assay ontology will
define individual SOP procedural components.
ACKNOWLEDGEMENTS
This research was funded as part of the EUMODIC project
(funded by the European Commission under contract number LSHG-CT-2006-037188). The authors also thank the
UK Medical Research Council for support.
REFERENCES
1.
2.
3.
4.
5.
6.
4
Hancock, J.M. and A.M. Mallon, Phenobabelomics--mouse phenotype data resources. Brief Funct
Genomic Proteomic, 2007. 6(4): p. 292-301.
Smith, C.L., C.A. Goldsmith, and J.T. Eppig, The
Mammalian Phenotype Ontology as a tool for annotating, analyzing and comparing phenotypic information. Genome Biol, 2005. 6(1): p. R7.
Gkoutos, G.V., et al., Using ontologies to describe
mouse phenotypes. Genome Biol, 2005. 6(1): p.
R8.
Knowlton, M.N., et al., A PATO-compliant zebrafish screening database (MODB): management of
morpholino knockdown screen information. BMC
Bioinformatics, 2008. 9: p. 7.
Green, E.C.J., et al., EMPReSS: European Mouse
Phenotyping Resource for Standardized Screens.
Bioinformatics, 2005. 21(12): p. 2930-2931.
Mallon, A.M., A. Blake, and J.M. Hancock, EuroPhenome and EMPReSS: online mouse phenotyp-
ing resource. Nucleic Acids Res, 2008.
36(Database issue): p. D715-8.
Taylor, C., et al., Promoting Coherent Minimum
Reporting Requirements for Biological and Biomedical Investigations: The MIBBI Project. Nat.
Biotechnol., 2008.
Developing an application focused experimental factor ontology:
embracing the OBO Community
James Malone*, Tim F. Rayner, Xiangqun Zheng Bradley and Helen Parkinson
EBI, Wellcome Trust Genome Campus, Hinxton, Cambridge, CB10 1SD, UK
ABSTRACT
Motivation: The recent crop of bio-medical standards has
promoted the use of ontologies for describing data and for
use in database applications. The standards compliant
ArrayExpress database contains data from >200 species
and >110,000 samples used in genotyping, gene expression
and other functional genomics experiments. We considered
two possible approaches in employing ontologies in
ArrayExpress: select as many ontologies as cover the
species, technology and sample diversity, choosing where
there are non-orthogonal resources and attempt to make
them interoperable; or build an extensible interoperable
application ontology. Here we describe the development of
an application focused Experimental Factor Ontology and
describe its use at ArrayExpress.
www.ebi.ac.uk/ontology-lookup/browse.do?ontName=EFO
1
INTRODUCTION
The value of having explicit and rich semantic
representations of data is becoming increasingly clear in
bioinformatics. This is apparent in the emergence of the
OBO foundry (Smith et al., 2007) and numerous metadata
standards (http://www.mibbi.sf.net). The OBO foundry
promotes the development of orthogonal ontologies that are
expressed in a common shared syntax, use unique
namespace identifiers and explicit textual definitions for all
ontology terms. These ontologies give us the terminology to
describe the level of detail that content standards such as
MIAME require. Underpinning this increased focus on the
use of ontologies is that richer and explicit representations
enhance interoperability and facilitate machine readability.
As the numbers of ontologies and standards increase, the
complexity of supporting standards using ontologies also
increases.
In this paper we describe development of the
Experimental Factor Ontology (EFO), an application
focused ontology. EFO models the experimental variables
(e.g. disease state, anatomy) based on an analysis of such
variables used in the ArrayExpress database. The ontology
has been developed to increase the richness of the
*
To whom correspondence should be addressed.
Email: [email protected]
1
annotations that are currently made in the ArrayExpress
repository, to promote consistent annotation, to facilitate
automatic annotation and to integrate external data. The
methodology employed in the development of EFO involves
construction of mappings to multiple existing domain
specific ontologies, such as the Disease Ontology (Dyck and
Chisholm, 2003) and Cell Type Ontology (Bard et al, 2005).
This is achieved using a combination of automated and
manual curation steps and the use of a phonetic matching
algorithm. This mapping strategy allows us to support the
needs of various ArrayExpress user groups who
preferentially use different ontologies, to validate existing
ontologies for coverage of real world high throughput data
in public repositories and to provide feedback to the
developers of existing ontologies. An additional reason to
have a local application ontology – rather than simply create
an enormous cross product ontology (i.e. classes created by
combining multiple classes from other ontologies)– is that
the structure of such an ontology may be challenging for
many users and time consuming to produce (Bard and Rhee,
2004). Instead, data acquisition tools can employ one
ontology rather than many external ontologies.
Brinkley et al. (2006) highlight the potential value in
reference ontologies for performing mapping and
integration for building application ontologies. However, at
present these frameworks and all necessary reference
ontologies do not exist. We therefore exploit the use of the
several OBO Foundry ontologies as reference ontologies in
contrast to the definition discussed by Brinkley et al by
employing a softer and more cautious view of these
ontologies. Specifically, we aim to map to the concept
names and definitions provided by external ontologies
without importing covering axioms, thereby reducing the
potential for conflict and removing an obstacle for
interoperability. Instead we use references in the same way
many OBO Foundry ontologies reference external resources
using a pointer to their identifier. This strategy avoids
‘bedroom ontology development’ wherein ontologies are
developed ab initio without considering the reuse of existing
ontologies. By re-using and mapping we leverage the user
supplied annotations and existing ontologies.
The EFO is represented in the web ontology language
(OWL) thereby conforming to an accepted common
representation and we also implement a policy of unique
J. Malone et al.
namespace identifiers and definitions for all terms, as
encouraged by OBO. Finally, we assess our ontology posthoc using semi-automated methods to assess the coverage
we have obtained in terms of our set of use cases (described
in our web resource http://www.ebi.ac.uk/microarraysrv/efo/index.html) and, hence, assess the suitability of the
ontology for the task at hand.
2
METHODOLOGY
Since the EFO is an application ontology, we developed a
well defined set of requirements based on our needs for
annotating experimental data. ArrayExpress typically has ~
five annotations per biological sample, and the most
important annotations are those that contain information on
the experimental variables. These are both biological i.e.
properties of the experimental samples (e.g. sex or
anatomy), and procedural; properties of protocols used to
treat the samples (e.g. sampling time or treatment with
compound). The initial focus in developing the EFO is on
the former as they are more likely to be present in a
reference ontology (i.e. non-numeric) and can be
automatically discovered in unstructured data. This is an
important use case for ArrayExpress as thousands of
experiments are imported from the Gene Expression
Omnibus where the sample annotation is essentially uncurated free text. Additionally from analysis of user queries,
biological information is more commonly queried than
procedural information.
Figure 1 High level classes from the EFO
2.1
Mapping, curating and integrating
Our approach is a middle-out ontology methodology as
described by Uschold and Gruninger (1996). In this
method, we start with a core of basic terms identified from
our use cases and specialize and generalize as required. Our
set of initial core terms already provided some structure as
the more specific concepts (called factor values) were
grouped into factor categories. We then created generalized
classes to give some additional structure to our ontology
(shown in Figure 1). The structure at the highest level has
been designed to be simple, and intuitive to biologists and
2
the curators, who will be the primary users of the EFO in the
short term, by constructing this as an abstraction of the
existing structure in ArrayExpress.
EFO terms have no internal text definitions by design,
instead we leverage the mapping strategy defined below to
create links to text definitions created by domain experts.
The mapping strategy involves selecting likely reference
ontologies and evaluating their coverage of terms present in
ArrayExpress. This includes ontologies such as the Disease
Ontology which also has many mapped terms, the Cell Type
Ontology and Zebrafish Anatomy and Development
ontology (Sprague et al., 2006) and the NCI thesaurus
(Fragoso et al., 2004) which has human, mouse and rat
terms related to cancer (Figure 2).
Figure 2 The intersection of the EFO and reference ontologies
To perform our mapping and add terms to EFO we used
the following iterative methodology:
ƒ
Identify OBO Foundry ontologies relevant to an EFO
category based on annotation use cases
ƒ
Create subset of classes of relevance to the ontology,
e.g. classes under disease for disease ontology
ƒ
Perform mapping using text mining phonetic
matching algorithm. This produces a list of candidate
ontology class matches.
ƒ
Manually validate matched ontology classes and
curate where necessary
ƒ
Manually map high quality annotations (identified as
present in the ArrayExpress data warehouse) to
multiple source ontologies
ƒ
Consider number of instances of terms used in
ArrayExpress to determine depth and breadth
ƒ
Integrate into EFO, adding appropriate annotation
values to definition and external ontology ID
ƒ
Structure EFO to provide an intuitive hierarchy with
user friendly labels
2.2
Phonetic matching
Our matching approach uses the Metaphone (Phillips, 1990)
and Double Metaphone algorithms (Phillips 2000) which
were selected following an empirical study of commonly
used matching algorithms and their utility in the biomedical
Developing an application focused functional genomics ontology: embracing the OBO Community
domain. We were particularly interested in algorithms
yielding low false positive rates, as we wished to use the
same algorithm for automatic annotation of incoming data.
We matched the user supplied cell type terms deposited in
ArrayExpress with the Cell Type Ontology using Soundex
(http://en.wikipedia.org/wiki/Soundex), Levenshtein edit
distance (Levenshtein, 1966), Metaphone (Phillips, 1999)
and Double Metaphone (Phillips, 2000) algorithms.
Synonyms and term names were used during the matching
process and matches were either single or multiple. For the
purposes of automated annotation, single matches are
obviously more desirable. The Metaphone algorithm yielded
the lowest false positive rate, with 98% of the matches
mapping to single ontology terms, and of these only 6%
were deemed to be invalid following inspection by an expert
curator. However, the overall coverage of the input term list
was relatively low (17% of all terms matched). In
comparison, the Double Metaphone algorithm provided
higher list coverage (50% of terms) at the expense of
generating a smaller proportion of single matches (48% of
total matches) and a much higher false positive rate (34% of
single matches). The Levenshtein and Soundex algorithms
yielded results similar to the Metaphone and Double
Metaphone algorithms, respectively, but both generated
slightly higher levels of false positives. A combined strategy
was therefore implemented, using Metaphone for a first pass
and then falling back to Double Metaphone for those terms
not matched by Metaphone. Using this strategy with curator
supervision to select the correct term in the multiple-match
cases yielded the highest overall number of matches with
minimal human intervention. Verified matched terms
identified by this strategy were included in the EFO and
placed manually in the hierarchy.
2.3
resources, as indicated by the definition citation annotation
property.
As an early version, the ontology still has parts that are
under review and is evolving. In particular, the hierarchy
still contains classes that are likely to be moved and
changed to add more structure as it is relatively flat at
present. Furthermore, the additional group of use case
covering cross species queries, e.g. disease and mouse
model of disease, and the representation of anatomical parts
in different species are required but are currently not
supported by the EFO. However, as the iterative engineering
process is ongoing, these will be addressed in the near
future. Where possible we will use existing resources to
address these use cases.
3.1
Validation
The ArrayExpress data flow doubles on a yearly basis. This
allows us to constantly validate the ontology against fast
changing annotation with a variety of granularities. It also
allows us to develop the ontology against emerging use
cases. We have implemented an iterative evaluation of the
ontology against the data content of the ArrayExpress
repository, against newly submitted data for curation
purposes and also against the ArrayExpress data warehouse
– a set of additionally annotated and curated data which
represents the ArrayExpress ‘gold standard’. As the
ontology evolves it will be used daily by the ArrayExpress
production team and incremental versions will be tested
internally prior to public release. Early stage evaluation is
performed semi-automatically by mapping between the
ontology and very large meta-analyzed curated experiments
and by comparison with reference ontologies. We were able
Ontology conventions
Naming conventions described by Schober et al. (2007)
were used. Specifically, class labels are intended to be
meaningful to human readers, short and self-explanatory.
They are singular and conform to the conventional linguistic
and common usage of the term, for example, the term
Huntingdon’s disease has a capital H since it is a proper
noun, whereas cancer would not. Identifiers have the
format EFO:00000001, where a unique integer identifies a
term and EFO identifies the ontology. We use an alternative
term annotation property to capture synonyms for class
labels, text definitions are not provided at present. The
ontology is developed in Protégé and converted to OBO
format for display in OLS.
3
THE EFO
Part of the hierarchy visualized in OLS is shown in Figure
3. The current version of EFO has ~800 child terms of the
class experimental factor. The majority of these have been
mapped to external reference ontologies and knowledge
Figure 3 EFO term ‘neoplasia’ visualized in OLS
3
J. Malone et al.
to assess granularity and overall coverage of the ontology
and structure is manually evaluated by the curators who use
the ontology.
Version 0.1 of EFO produces automated mappings
comparable in coverage (~35%) for a 6000 sample test set
between the EFO and the NCI thesaurus. Replacing the NCI
thesaurus with the EFO reduced false positives and multiple
matches by an order of magnitude (60% reduced to 8.6%).
We believe by continuing an iterative process of mapping,
curating and integrating EFO terms alongside an iterative
evaluation strategy and restructuring the ontology we can
continue to improve the quality and coverage of the
ontology throughout its lifecycle.
4
DISCUSSION
It is our belief that application ontologies such as the EFO
should be constructed with a principal to minimize
redundancy and maximize information sharing. Wherever
possible, mapping to external resources such as OBO
Foundry ontologies increases interoperability through a
common and shared understanding. Furthermore, this
removes the temptation to ‘reinvent the wheel’ and allows
the exploitation of the efforts currently underway to
represent particular communities. It also permits updating
when reference ontologies change.
A complication of this approach is the implication of
mapping to external ontology concepts and their implicit
hierarchy. In EFO our ‘meaning’ is limited to the textual
definitions of the concepts externally mapped to EFO terms.
Importing and accepting all axioms associated with concepts
is a desirable long term goal. However the potential for
conflicting logical definitions and lack of an intuitive
standardized and easy to use upper ontology framework
have caused us to initially defer this task. BFO (Grenon et
al., 2004) was not considered as an upper level ontology for
EFO in its earliest form as the primary focus of this project
is the application of the ontology and rapid development.
However, mapping to BFO (or some other upper level
ontology) is something we are now beginning to look
into.for future development and will appear in the
forthcoming future releases.
The OBO Foundry has resolved issues, of orthogonal
coverage and unique namespace identifiers and has made
our task easier. In the future we will make bimonthly
releases of EFO, continue the validation process, consider
requests for new terms and map additional data resources to
the EFO. GEO data imported into the ArrayExpress
framework is already mapped during import, and any data
resource with biological annotation could be mapped semiautomatically. Obvious candidates include Uniprot and
other gene expression databases which are targets for
integration with ArrayExpress. Version 0.2 of EFO is
available from the EBI Ontology Lookup Service,
4
comments
and
questions
[email protected]
can
be
sent
to
ACKNOWLEDGEMENTS
The authors are funded in part by EC grants FELICS
(contract number 021902), EMERALD (project number
LSHG-CT-2006-037686), Gen2Phen (contract number
200754) and by EMBL. Thanks to the ArrayExpress
production team: Anna Farne, Ele Holloway, Margus Lukk,
and Eleanor Williams for useful comments.
REFERENCES
Bard, J, Rhee, SY et al. (2005) An ontology for cell types. Genome
Biol. 6(2): R21.
Bard, J, and Rhee, SY (2004) Ontologies in biology: design,
applications and future challenges. Nature Rev Gen 5, 213-222.
Brazma, A, Hingamp, P et al. (2001) Minimum information about
a microarray experiment (MIAME)—toward standards for
microarray data. Nature Genetics 29, 365-371.
Brinkley, JF, Suciu, D et al. (2006) A framework for using
reference ontologies as a foundation for the semantic web. In
Proceedings, American Medical Informatics Association Fall
Symposium, 96-100, Bethesda, MD.
Dyck, P and Chisholm, R (2003) Disease Ontology: Structuring
Medical Billing Codes for Medical Record Mining and Disease
Gene Association. Proceedings of the Sixth Annual Bioontologies Meeting, Brisbane, 2003, 53-55.
Fragoso, G, de Coronado, S, et al (2004) Overview and Utilization
of the NCI Thesaurus. Comp Func Gen 5:8:648-654.
Grenon, P., Smith, B. et al. (2004) Biodynamic ontology: applying
BFO in the biomedical domain. In Ontologies in Medicine (ed.
Pisanelli, D.M.) 20–38 (IOS, Amsterdam, 2004).
Levenshtein, VI (1966) Binary codes capable of correcting
deletions, insertions, and reversals. Soviet Physics Doklady
10:707–710.
Parkinson H, Kapushesky M, et al. (2006). ArrayExpress - a public
database of microarray experiments and gene expression
profiles. Nucl Acids Res 35, D747-750.
Phillips L (1990) Hanging on the Metaphone. Comp Lan 7: 39-49.
Phillips L (2000) The Double Metaphone Search Algorithm.
C/C++ Users Journal.
Schober, D, Kusnierczyk, W et al. (2007) Towards naming
conventions for use in controlled vocabulary and ontology
engineering, Proceedings of the Bio-Ontologies Workshop,
ISMB/ECCB, Vienna, July 20, 2007, 87-90.
Smith B, Ashburner, M, et al. (2007). The OBO Foundry: coodinated evolution of ontologies to support biomedical data integration. Nat Biotechnol 25, 1251.
Sprague, J, Bayraktaroglu L, (2006) The Zebrafish Information
Network: the zebrafish model organism database. Nucleic Acids
Res. 34, D581–D585.
Uschold, M, Grüninger, M (1996) Ontology: Principles, Methods
and Applications. Knowledge Engineering Review 11(2), 93155.
TGF-beta Signaling Proteins and the Protein Ontology
#1
1
1
2
1
1
Cecilia Arighi
, Hongfang3 Liu# , Darren Natale
, Winona Barker , Harold Drabkin , Zhangzhi Hu ,
2
1
Judith Blake , Barry Smith and Cathy Wu *
3
1
Georgetown University Medical Center, Washington DC, USA; 2The Jackson Laboratory, Bar Harbor, USA; State University of New York at Buffalo, Park Hall, Buffalo,USA.
ABSTRACT
Motivation: The Protein Ontology (PRO) addresses the
need for a formal description of proteins and their evolutionary relationships. PRO is authored via manual curation on the basis of content derived automatically from
various data sources. Curation is needed to ensure correct
representations of relationships both internally (between
PRO nodes) and externally (to other ontologies). Focusing
specifically on the TGF-beta signaling proteins, we describe how this ontology can be used for multiple purposes, including annotation, representation of objects in
pathways, data integration, and the representation of biological system dynamics and of disease etiology.
1
INTRODUCTION
The Open Biomedical Ontologies (OBO) Foundry is a
collaborative effort to establish a set of principles for ontology development with the goal of creating a suite of
orthogonal interoperable reference ontologies in the biomedical domain [Smith et al., 2007]. The Foundry ontologies are organized along two dimensions: (1) granularity (from molecule to population); and (2) relations to
time (objects, qualities, processes). In terms of this
scheme, PRO is a representation of entities on the level of
granularity of single molecules. It treats the molecules
themselves, and interoperates with other ontologies, like
the Sequence Ontology (SO) and the Gene Ontology
(GO), for protein qualities and processes. PRO encompasses (i) a sub-ontology of proteins based on evolutionary relatedness (ProEvo), and (ii) a sub-ontology of the
multiple protein forms produced from a given gene locus
(ProForm) [Natale et al., 2007]. Here we summarize the
current PRO framework focusing on the representation of
proteins from the TGF-beta signaling pathway since they
provide a rich body of protein annotation relating to a
wide spectrum of protein forms (derived from cleavage
and/or post-translational modifications (PTMs), alternative splicing, and sequence variants that are related to
disease).
*
#
To whom correspondence should be addressed.
Equal contribution to this work
2
THE PRO FRAMEWORK
Fig.1A shows the current working model and a subset of
the possible connections to other ontologies. The root in
the ontology is the class protein, which is defined as a
biological macromolecule that is composed of amino acids linked in a linear sequence (a polypeptide chain), and
is genetically encoded. PRO terms are connected by the
relationship is_a or derives_from, both defined in the
OBO Relations Ontology [Smith et al., 2005].
ProEvo: Proteins with similar domain architecture (that
is, the same combination of domains in the same order)
and full-length sequence similarity are said to be homeomorphic; they share a common ancestor and, usually, a
specific biological function. Also, within any given homeomorphic group, there may be monophyletic subgroups of proteins that have distinct functions [Wu et al.,
2004a] [Mi et al., 2006]. ProEvo defines protein classes
based on these concepts, and captures the relationship
between these classes. An illustrative example in PRO is
depicted in Fig.1B-C. The PRO term PRO:000000008
TGF-beta-like cysteine-knot cytokine is defined as a protein with a signal peptide, a variable propeptide region
and a cysteine-knot domain (definition in Fig.1C). The
class represented by this term has seven children
(Fig.1B), each of which can be defined as a separate
group on the basis of a distinctive functional feature. PRO
represents the proteins and not the individual domains.
Thus, domain information is included in the ontology as
part of the annotation of ProEvo terms with a link to the
Pfam domain database [Finn et al., 2006] to indicate that a
given protein class has_part some domain (see Pfam annotation in PRO:000000008 in Fig.1C). Therefore, tracing the relation between different ProEvo nodes would
involve reasoning over the presence of a given domain.
The gene product class, which is the leaf node of ProEvo,
defines all protein products of strictly orthologous genes.
ProForm: This part of the ontology (Fig1A-B) describes
the subset of the translational products that is experimentally characterized, and includes definition of sequence
forms arising from allelic, splice, and translational variation, and from PTM and cleavage. Moreover, it allows
representation of proteins that are products of a gene fusion due to chromosomal translocation, such as
1
C. Arighi et al.
PRO:000000091 creb-binding protein/zinc finger protein
HRX that is encoded by part of the CREBBP gene at the
N-terminus and part of the MYST4 gene at the Cterminus. This form is observed in some cases of acute
myelogenous leukemia. In ProForm, equivalent protein
forms in different species are represented as a single node.
We use the derives_from relationship to describe the relation between a modified form and the parent protein. In
addition, each form is connected to other ontologies,
which provide the annotation (Fig.1A, C).
3
A PRO
Pfam
protein
Root Level
is_a
has_part
GO
• Derivation: common ancestor
• Source: PIRSF family
translation product of an evolutionarily-related gene
is_a
Gene-Level Distinction
translation product of a specific gene
• Derivation: specific allele or splice variant
• Source: UniProtKB
• Derived from post-translational modification
• Source: UniProtKB
Gene Level
Family Level
Root Level
participates_in
cellular component
translation product of a specific mRNA
part_of (complexes)
located_in (compartments)
derives_from
Modification-Level Distinction
ProEvo
has_function
is_a
Sequence-Level Distinction
Sequence Level
ProEvo
molecular function
biological process
• Derivation: specific gene
• Sources: PIRSF subfamily, Panther subfamily
ProForm
B
protein domain
Family-Level Distinction
Modification Level
BUILDING THE ONTOLOGY
For the current release (1.0) we focus on the set of proteins in the TGF-beta signaling pathway from the KEGG
pathway database [Kanehisa et al., 2002], which includes
the TGF-beta, the bone morphogenetic protein and activin-mediated signaling pathways. The ontology consists
of a total of 667 PRO terms, including 111 ProEvo and
544 ProForm terms. It covers 79 human/mouse orthologous proteins that mapped to 34 PIRSF homeomorphic
cleaved/modified translation product
Example:
TGF-beta receptor phosphorylated smad2 isoform1
is a phosphorylated smad2 isoform1
derives_from smad2 isoform 1
is a smad2
is a TGF-β receptor-regulated smad
is a smad
is a protein
MIM
ProForm
disease
agent_in
SO
sequence change
has_agent
PSI-MOD
protein modification
has_modification
C
PRO:000000001 protein
is_a PRO:000000008 TGF-beta-like cysteine-knot cytokine
def: A protein with a core domain composition consisting of a signal peptide, a variable propeptide region and a transforming growth factor beta like
domain, that is a cysteine-knot domain containing four conserved beta strands, S1-S4, which form two antiparallel beta sheets (S1-S2 and S3-S4)
interconnected by three disulfide bridges in a knot-like topology.
Pfam:PF00019 "has_part Transforming growth factor beta like domain".
is_a PRO:000000046 TGF-beta { PIRSF001787 }
def: A TGFB-like cysteine-knot cytokine whose propeptide region is a latent associated peptide (LAP) that remains associated to the mature TGF-beta
after cleavage and secretion, rendering TGF-beta inactive. This is the founding member of the cysteine-knot cytokine family, and is related to the
activin/inhibin, anti-Muellerian hormone, and bone morphogenic protein families, which are all involved in the regulation of cell growth and
differentiation.
is_a PRO:000000182 TGF-beta 1 (TGFB1 gene products) { PANTHER:PTHR11848:SF32 }
def: A TGF-beta that is a translation product of the TGFB1 gene.
is_a PRO:000000397 TGF-beta 1 isoform 1 (precursor) { UniProtKB: P01137, P04202 }
def: A TGF-beta 1 that is a translation product of a processed transcript of the TGFB1 gene, and includes all core domains (signal peptide, latent
associated peptide and a transforming growth factor beta like domain), as in the human sequence UniProtKB:P01137. This form is a precursor.
GO:0005796 "located_in Golgi lumen [PMID:1544940, TaxID:9606]".
derives_from PRO:000000512 TGF-beta 1 isoform 1 cleaved and glycosylated
MOD:00693 "has_modification glycosylated residue".
is_a PRO:000000617 TGF-beta 1 isoform 1 cleaved and glycosylated 1 (TGF-beta 1 latent peptide)
is_a PRO:000000618 TGF-beta 1 isoform 1 cleaved and glycosylated 2 (Latent complex) { P01137 (30-279, N-glycosylated Asn-82):(279-390) }
GO:0005160 "NOT has_function transforming growth factor beta receptor binding [PMID:3162913, TaxID:9606]".
GO:0005578 "located_in proteinaceous extracellular matrix [PMID:9008713, TaxID:10090]".
derives_from PRO:000000513 TGF-beta 1 isoform 1 cleaved form
is_a PRO:000000616 TGF-beta 1 isoform 1 cleaved 1 (TGF-beta 1 active peptide) { P01137-1 (279-390) }
def: A TGF-beta 1 isoform 1 cleaved form that has been processed by proteolytic cleavage and released from the latent complex. This form is
the active mature peptide.
GO:0005160 "has_function transforming growth factor beta receptor binding [PMID:14764882, TaxID:10090]".
GO:0005578 "located_in proteinaceous extracellular matrix [PMID:9008713, TaxID:10090]".
is_a PRO:000000401 TGF-beta 1 sequence variant 4 (precursor) { His-218 in human sequence, VAR_017607 }
def: A TGF-beta 1 that is a translation product of the a polymorphic sequence variant of TGFB1 gene that has a His residue at the position
equivalent to Arg-218 in the human sequence UniProtKB:P01137.
MIM:131300 "agent_in CAMURATI-ENGELMANN DISEASE [PMID:12493741, TaxID:9606]".
SO:1000093 "has_agent mutation_causing_amino_acid_substitution".
derives_from PRO:000000516 TGF-beta 1 sequence variant 4 cleaved form
is_a PRO:000000627 TGF-beta 1 sequence variant 4 cleaved 1 (TGF-beta 1 active peptide) { VAR_017607 (279-390) }
GO:0008083 "has_function (increased) growth factor activity [PMID:12843182, TaxID:9606]".
is_a PRO:000000628 TGF-beta 1 sequence variant 4 cleaved 1 (TGF-beta 1 latent peptide) { VAR_017607 (30-279) }
SO:1000093 "has_agent mutation_causing_amino_acid_substitution".
SO:1000124 "has_agent mutation_causing_partial_loss_of_function_of_polypeptide [PMID:11278244, TaxID:9606]".
Fig.1- PRO framework and DAG view of the ontology. A) Current working model and a subset of the possible connections to other ontologies. B) Snapshot of the ontology (partial view) in OBO Edit 1.1 including terms representing ProEvo and ProForm. C) A PRO example illustrated by the TGF-beta 1 protein. The above is a partial view, not all forms are listed, and only key annotations are shown.
2
C. Arighi et al.
families and 36 Pfam domains. An automated process has
been developed to generate PRO nodes from PIRSF [Wu
et al., 2004a] and iProClass [Wu et al., 2004b] databases,
UniProtKB [UniProt Consortium, 2008], as well as MGI,
Pfam, and PANTHER [Mi et al, 2005]. The computationally-generated file is in OBO format and we use OBO
Edit 1.1 [Day-Richter et al., 2007] as the curation platform. Manual curation includes (i) merging of nodes, for
example whenever PIRSF and PANTHER families represent the same homeomorphic protein class (same membership), (ii) reviewing the literature and sequence analysis to verify or create the protein forms, e.g., analyzing
what combination of modifications occurs in a specific
form, and determining what forms are equivalent in
mouse and human. Furthermore, new ProForm nodes can
be created for newly characterized isoforms or sequence
variants not yet represented in UniProtKB (e.g.,
PRO:000000478 smad5 isoform 2, and PRO:000000483
smad9 isoform 2). Names of ProEvo nodes are adapted
from the underlying data sources or from the literature.
The names of ProForm nodes are based on their parent
node (Fig.1B, under the ProForm bracket). All PRO terms
have a definition that conforms to OBO foundry standards
with examples delineated in Fig.1C. A reference to conserved motifs and domain regions is used whenever possible and, in some cases, examples are supplied. Each
PRO definition has source attribution to PubMed ID, PRO
curator, or other resource ID. In addition, the annotations
are introduced via cross-references to other ontologies.
Currently, the majority of model organism databases supply GO annotation to a gene object rather than to a specific protein form. PRO assigns the GO terms to the specific forms (Fig.1C). In the example above, although
TGF-beta 1 is annotated in databases with the
GO:0005160 transforming growth factor beta receptor
binding, this term is not appropriate to annotate the precursor (PRO:000000397), but rather the active peptide
(PRO:000000616). So the advantage of the PRO framework is that it can provide a basis for more accurate annotation. To further illustrate the importance of this statement, Fig.1C shows some of the nodes and relationships
for the TGF-beta 1 protein, thereby demonstrating the
complexity and variety of sequence forms that can be
derived from a given parent sequence. TGF-beta 1 precursor is a dimer and undergoes two cleavages−by a signal
peptidase and by furin in the Golgi−to generate two functionally important chains: the TGF-beta 1 mature and the
latent peptide (PRO:000000617). These two chains remain associated (as a latent complex) until proteases in
the extracellular space degrade the latent peptide. The
latent complex is represented by PRO:000000618,
whereas the active mature protein is represented by
PRO:000000616. Note that only the latter is associated
with the GO term corresponding to receptor binding activ-
ity, and that the TGF-beta 1 isoform 1 precursor and any
of its derived forms differ in cellular localization. In addition, an arginine to histidine variant (R218H) in the human protein is responsible for the Camurati-Engelmann
disease (agent_in the disease). This mutation affects the
stability and conformation of the latent peptide, elevating
the levels of free (active) mature peptide. This situation is
formally represented in PRO by associating the corresponding SO terms to the corresponding products (see
PRO:000000401 and its children nodes) and, also, by
adding a modifier to the has_function relationship to reflect the constitutively active mature peptide
(PRO:000000627). Also note that there is no term corresponding to the latent complex derived form for this variant. A total of 1667 annotations have been added to PRO
nodes in release 1.0. Table 1 shows the statistics for GO
terms. The examples illustrate how appropriate annotation
can be assigned to appropriate protein forms.
Table 1: Statistics on GO terms in PRO release 1.0
GO term
Molecular
Function
OBO Relation
has_function
#
terms
181
NOT
has_function
part_of
43
NOT
part_of
5
Cellular
Component
located_in
171
Biological
Process
participates_in
235
Cellular
Component
Complex
4
38
Example
PRO:000000650
smad 5 isoform 1
phosphorylated 1
PRO:000000478
smad 5 isoform 2
PRO:000000178
RING-box protein
2 isoform 1
PRO:000000179
RING-box protein
2 isoform 2
PRO:000000457
noggin isoform 1
cleaved 1
PRO:000000086
chordin isoform 1
GO:0046332
SMAD binding
GO:0046332
SMAD binding
GO:0000151
ubiquitin ligase
complex
GO:0000151
ubiquitin ligase
complex
GO:0005615
extracellular
space
GO:0001501
skeletal development
DISSEMINATION OF PRO
The results of the PRO project are disseminated through
several mechanisms: the entire ontology and associated
wiki are both accessible through the PRO public website
(http://pir.georgetown.edu/pro/), as well as through the
OBO Foundry (http://www.obofoundry.org/), and the
National Center for Biomedical Ontology (NCBO) BioPortal (http://www.bioontology.org/bioportal.html).
5
PRO AND ITS USER COMMUNITY
Any project that needs to specify protein objects of the
type described in the ontology can benefit from PRO. The
TGF-beta signaling pathway described above shows how
the protein ontology can assist in the explicit annotation
of states of a molecule. These states are natural components of pathway ontologies or databases such as INOH
Event Ontology [Kushida et al., 2006] or Reactome [Vastrik et al., 2007] (the latter does contain the relevant entities, but as accessions only; they are not as yet formed
3
C. Arighi et al.
into an ontology structure which supports reasoning). As
biomedical data expand, it will be increasingly important
to explicitly represent these protein forms so that representations of attributes can be attached to the appropriate
entities. Fig.2 shows part of the TGF-signaling pathway
as described by Reactome with PRO terms mapped to the
associated Reactome events. This improves the mapping
of the entities involved in the pathway, and gives a more
accurate and complete framework for researchers to analyze their data. In addition, PRO allows modeling of the
specific objects involved in a given disease, as the TGFbeta 1 sequence variant 4 case described above (Fig.1C).
Other examples include the representation of (1) a cleaved
form
of
rho-associated
protein
kinase
1
(PRO:000000563), which is constitutively active in the
mouse myopathy model and in human heart failure patients, and of (2) smad4 and BMP receptor type-1A sequence variants associated with a common disease
allowing the inference of a specific pathway failure. PRO
terms could also be adopted by GO to accurately define
protein complexes in the cellular component ontology.
PRO could potentially be used for cross-species comparison of protein forms, since only the forms with experimental evidence are included (with the associated literature and taxon IDs). Finally, PRO could be adopted where
data integration at the molecular level of proteins is
needed, as in systems biology or in translational medicine.
PRO:000000618;
PRO:000000552
Large latent complex
PRO:000000616
Activated TGF-beta1
REACT_6888.1
1
PRO:000000616;
PRO:000000409;
PRO:000000523;
PRO:000000313;
PRO:000000468;
Recruitment of SARA
and Smad2
PRO:000000616;
PRO:000000409;
PRO:000000523;
Receptor I
phosphorylation
2
REACT_6872.1
P
3
REACT_6945.1
PRO:000000409;
PRO:000000616
Activated TGFbeta1 bound to
receptor II
5
PRO:000000616;
PRO:000000409;
PRO:000000522
Recruitment of
receptor type I
PRO:000000468
Funding for PRO development is provided by NIH grant 1 R01
GM080646-01.
REFERENCES
P SARA
Smad 2
Smad 2
REACT_6816.1
ACKNOWLEDGEMENTS
PRO:000000616;
PRO:000000409;
PRO:000000523;
PRO:000000313;
PRO:000000650
Smad2 phosphorylation
P SARA
4
pathway. The significance of the framework can be summarized as follows: (1) it provides a structure to support
formal, computer-based inferences based on data pertaining to shared attributes among homologous proteins; (2) it
helps us to delineate the multiple protein forms of a gene
locus; (3) it provides important interconnections between
existing OBO Foundry ontologies; (4) it provides a
framework that can be adopted by other ontologies and/or
databases, as for example, to better define objects in
pathways, or complexes or in disease modeling; (5) it
allows the community to annotate their proteins of interest. Finally, it offers a comprehensive picture of the protein realm by connecting protein evolution, function,
modification, variants, gene ontology and disease.
P
6
REACT_6923.1
REACT_6979.1
Smad 2
7
PRO:000000366
Smad 2
PRO:000000650
P
Smad 4
REACT_6760.1
8
Smad 2
Smad2/Smad4 complex
binding to other
transcription factors to start
transcription
P
Smad 4
Smad 2
Transc factors
REACT_6726.1
Smad 2
P
Smad 4
PRO:000000650;
PRO:000000366
P 10
Smad 4
9
Nucleus
Cytoplasm
PRO:000000650;
PRO:000000366
R-Smad/CoSmad complex
Fig.2- PRO and the Reactome TGF-beta signaling pathway
(REACT_6844). Each step in the pathway is described by a
Reactome event ID. Bold PRO IDs indicate objects that undergo
some modification that is relevant for function (the modified
form is underlined).
6
CONCLUSION
We illustrated key aspects of the PRO framework through
reference to proteins involved in the TGF-beta signaling
4
Day-Richter J., Harris M.A., Haendel M.; Gene Ontology OBOEdit Working Group, Lewis S. (2007) OBO-Edit--an ontology
editor for biologists.Bioinformatics, 23:2198-2200.
Finn R.D., Mistry J., Schuster-Bockler B., Griffiths-Jones S.,
Hollich V., et al. (2006) Pfam: clans, web tools and services.
Nucleic Acids Res., 34:D247-251.
Vastrik I., D'Eustachio P., Schmidt E., Joshi-Tope G., Gopinath
G., et al. (2007) Reactome: a knowledge base of biologic
pathways and processes. Genome Biology, 8:R39.
Kanehisa M, Goto S, Kawashima S, Nakaya A. (2002) The
KEGG databases at GenomeNet. Nucleic Acids Res. 30:42-46.
Kushida T., Takagi T., Fukuda K.I. (2006) Event Ontology: A
pathway-centric ontology for biological processes. Pacific
Symposium on Biocomputing 2006, 11:152-163.
Mi H, Lazareva-Ulitsky B, Loo R, Kejariwal A, Vandergriff J,
et al. (2005) The PANTHER database of protein families,
subfamilies, functions and pathways. Nucleic Acids Res.,
33:D284-288.
Natale D., Arighi C., Barker W.C., Blake J., Chang T., et al.
(2007) Framework for a Protein Ontology. BMC Bioinformatics, 8(Suppl 9):S1.
Smith B., Ceuster W., Klagger B., Kohler J., Kumar A., et. al.
(2005) Relations in Biomedical Ontologies. Genome Biol.,
6:R46.
Smith B., Ashburner M., Rosse C., Bard J., Bug W., et al.
(2007) The OBO Foundry: coordinated evolution of ontologies to support biomedical data integration. Nat Biotechnol.,
25: 1251-1255.
UniProt Consortium (2008) The Universal Protein Resource
(UniProt). Nucleic Acids Res., 36(Database issue):D190-195.
Wu C.H., Nikolskaya A., Huang H., Yeh L-S., Natale D.A., et
al. (2004a) PIRSF family classification system at the Protein
Information Resource. Nucleic Acids Res., 32:D112-114.
Wu C.H., Huang H., Nikolskaya A., Hu Z., Yeh L.S., et al.
(2004b) The iProClass Integrated database for protein functional analysis. Computational Biol. and Chem, 28: 87-96.
Tamoxifen to Systemic Lupus Erythematosus: Constructing a Semantic Infrastructure to
Enable Mechanism-based Reasoning and Inference from Drugs to Diseases
Angela Xiaoyan Qu1,3,4*, Ranga C. Gudivada1,3, Eric K. Neumann5 , Bruce J Aronow1,2,3*
1
2
3
University Departments of Biomedical Engineering and Pediatrics , University of Cincinnati and Division of Biomedical Informatics , Cincinnati
4
5
Childrens Hospital Medical Center; Procter & Gamble Pharmaceutical , Cincinnati OH; Clinical Semantics Group , Lexington, MA
ABSTRACT
Discovering new clinical applications for existing drug products
and predicting novel drug combinations for improved efficacy
represent promising opportunities for both pharmaceutical
development and personalized medicine. To enable these
efforts, we have sought to develop a systematic framework for
representation and inference of drug/disease relationships
based on mechanistic knowledge. To do this we have
developed a Disease-Drug Correlation Ontology (DDCO) that
provides a framework for asserting entity and relationship type
characteristics of heterogeneous data from pharmacological,
medical, and genetic, and other biological domains. The
DDCO, formalized in OWL, allows for the representation of
multiple ontologies, controlled vocabularies, and data
schemas and normalized mappings of relationships between
elements of each source. In the present study we used the
DDCO framework to form relationships across a collection of
data sources including DrugBank, EntrezGene, OMIM, Gene
Ontology, SNOMED, MeSH Anatomy, and other sources in
UMLS, to construct an extensible Pharmacome-GenomeDiseasome network. As an example, we illustrate the utility of
this approach to simultaneously model biological processes
associated with disease processes, phenotypic attributes, and
mechanisms of drug action to predict a new indication for
Tamoxifen could be to treat the therapeutically challenging
disease entity Systemic Lupus Erythematosus.
1. INTRODUCTION
Drug repositioning, i.e. the use of established drugs to treat
diseases that are not established as indications for its use,
represents a promising avenue based on its lower development
cost and availability of extensive data and knowledge from
prior research1. Despite impressive successes shown by
repositioned drugs, most of these are the result of “serendipity”
–ie unexpected findings made during or after late phases of
clinical study. Thus, a forecasting model that could improve
data capture, integration, analysis, and prediction of potential
new therapeutic indications for drugs based on integrated
biomedical knowledge around drug and disease mechanisms is
highly desirable. Currently, most drug-oriented databases e.g.
PharmGKB2, KEGG3 and DrugBank4 tend to support limited
dimensionality of mechanism-associated relationships and lack
the multi-disease gene and phenotype relationships that are
likely to be necessary to infer between disparate diseases.
*
Corresponding authors: [email protected], [email protected].
The advancements of Semantic Web (SW)5 and related
knowledge representation technologies provide a promising
platform for semantic integration of heterogeneous data and
knowledge interoperability. Hypothesizing that associating
comprehensive biomedical information and prior knowledge
around pharmacological entities (i.e. biological, chemical, and
clinical processes) and using SW principles and technologies can
facilitate reveal new knowledge such as novel indications for
known or unknown drugs, we devised a knowledge framework,
Disease-Drug Correlation Ontology (DDCO), using Web
Ontology Language (OWL) representation formalism to facilitate
mapping and assertion of these relationships across multiple
ontologies and hierarchically organized data sources.
The working ontology, DDCO, is thus an aggregation of
manual curation and integration of relevant components from
multiple existing ontologies, vocabularies, and database schemas.
We used the DDCO to link DrugBank, OMIM, EntrezGene,
KEGG, BioCarta, Reactome, and UMLS; and the data was
semantically integrated into a Resource Description Framework
(RDF) network. Using this, we present as an example scenario the
implication of Tamoxifen, an established drug product, as
potential therapeutic for systemic lupus erythematosus (SLE).
We propose that further populating knowledge bases with similar
structure will enable both new indications and the identification
of synergistic drug combinations.
2. DRUG-DISEASE CORRLEATION ONTOLOGY
CONSTRUCTION
2.1 Ontology Development
Our goal is to devise a drug- and disease-centric knowledge
framework that serves both data integration and knowledge
exploitation needs. The ontology was designed with high-level of
granularity and aims to reuse knowledge components whenever
possible. Therefore, the first step for our ontology development
effort was to examine and select from previously existing
resources that allow efficient knowledge mapping and sharing
among independent data sources. We used UMLS Semantic
Network7 to construct the scaffold of the DDCO. Though
containing a set of broad semantic types and relationships
defining biomedical concepts, UMLS Semantic Type has
knowledge “gaps” and is insufficiently organized. For example,
for pharmacological domain, it only contains one Semantic Type,
“Pharmacologic_Substance” with only one child term,
1
Qu. AX. et al.
“Antibiotic”, which is far from the full representation of drugcentric entities. To fill in such gaps, we proposed to also use
below ontology or vocabulary sources (Fig.1):
• MeSH (medical subject headings): the controlled
vocabulary thesaurus in biomedical fields.
• NCI Thesaurus8: an ontology-like vocabulary in cancercentric disease areas
• The Anatomical Therapeutic Chemical Classification9: a
WHO
recommended
classification
system
for
internationally applicable methods for drug utilization
research
• Common Terminology Criteria for Adverse Event
(CTCAE) 10: A descriptive terminology and grade scales
for drug adverse event reporting
• Gene Ontology11
• SNOMED CT12: clinical health care terminology and
infrastructure
Ontology editor Protégé was used as the primary tool for
implementing the OWL framework13. To enhance the editing
and visualization flexibility, several plug-ins were also used
including PROMPT for ontology comparison and merging and
OwlViz for visualization. Ontology mapping and aligning
techniques were applied for concept and relationship
integration. In addition, manual modifications, such as pruning
irrelevant or duplicate branches or adding new
concepts/relationships, were performed to maximize integration
and minimize non-connectivity.
In addition, concept
restrictions and property constraints were also manually curated
to support the inferential capability enabled by description
logic. We used RACER14, a description logic reasoning system
with support for T-Box and A-box reasoning, to pose DL
queries for the ontology evaluation. On average, the
subsumption computations were completed within ten seconds
and we sought to solve any inconsistencies to assure the
integrity of the DDCO.
2.2
Ontology Model Metrics
While efforts to expand and refine the conceptualization are
continuing, the current DDCO contains 2046 classes (excluding
GO which was imported directly), with average sibling number
of 17 (maximum 35 and minimum 1) per class. There are total
of 221 properties, with 99 properties domain-specified, 69
range-specified, and 36 inverse-specified. These properties
include 135 selected UMLS Semantic Network relations, 40
SNOMED attributes, and 46 custom-defined properties for
constraint development. To further refine the entities, we
created 67 restrictions: 7 existential, 36 universal, and 25
cardinality. Fig.2 presents a top-level view of the ontology
concepts as well as properties connecting them. Fig.3 shows the
semantic model for the “Drug” entity including our curation of
concept restrictions including necessary and sufficient
restrictions. For example, one of the criteria to define a drug is it
needs to have at least one “active” ingredient (Fig.3).
2
Fig 1. Schematic view of Drug-Disease Correlation Ontology model
and the major resources (in oval) included
Fig 2. Top-level view of conceptual frames and domain/range
relationships in DDCO.
3.
APPLICATIONS
3.1. An Integrated Pharmacome-Genome-Diseasome
RDF Network
A key benefit of the Semantic Web is its ability to integrate
relevant data from different origins and in incompatible
formats. We used the DDCO as the knowledge framework to
integrate a diverse collection of data sources across
pharmacome- , genome-, and diseasome- domains.
Specifically, following data sources were used for extracting
and integrating relevant data:
• Pharmacome Data Source:
Drug-associated information was compiled from DrugBank.
The resultant data set contains 4,763 drug entries, including
over 1,400 FDA-approved drugs.
Tamoxifen to Systemic Lupus Erythematosus: Constructing a Semantic Infrastructure to Enable Mechanism-based Reasoning and
Inference from Drugs to Diseases
defined in the DDCO for each of our pharmacome, genome,
and diseasome domains. These models provide the required
mapping mechanism from the instance data to DDCO in order
to semantically annotate and relate different –omic entities.
The data parsed and extracted from above sources was
converted from various formats (i.e. RDF/OWL, XML, txt)
into RDF triples using different RDF converters in compliance
with the definitions by the DDCO model. The converted RDF
triples were further converted into N-Triple format using
Oracle RDF loaders before loading to the Oracle 10g release 2
RDF store16. With the assigned unique name space and the
shared identifiers, the data loaded in the RDF model are
thereafter integrated automatically in a seamless manner.
3.2 Exploiting Drug-Disease Association: A Scenario
from Tamoxifen to SLE
Fig 3. Partial view of properties and restrictions for Drug entity
modeling in DDCO
• Genome Data Source:
The annotation of human genes and interactome data including
BIND, BioGRID and HPRD data were downloaded was from
NCBI ftp site. Gene-pathway annotations were compiled from
KEGG, BioCarta, BioCyc and Reactome databases. The total
data set contains 15068 human genes annotated with 7124
unique GO terms, and 14899 gene-pathway associations.
•
Diseasome Data Source:
OMIM15 records were downloaded in XML format. OMIM ID
and the corresponding gene associations were downloaded
from NCBI EntrezGene ftp site.
To explore the implicit associations between drug and
disease, we need to understand the “explicit” relationships
between them too. Thus, we have extracted the known drugdisease associations (i.e. indication for FDA-approved drugs)
using UMLS 2007AC files from UMLS Knowledge Server.
We used the table MRCONSO.RRF to map the FDA-approved
drugs to the UMLS unique concept identifier (CUI). Next, the
table MRREL.RRF was used to extract the associated
indications for these CUI concepts. The semantic relationships
of “may_treat” and “may_be_treated_by” were used to restrict
the relationship mapping. To further refine the extraction and
eliminate false positive mapping, the semantic type “Chemicals
& Drugs” and “Disorders” were used to constrain the
association concepts. As a result, a total of 230,114 drugdisease associations were extracted.
As the next step to build the integrated RDF data graph, we
created models based on the logic and semantic relationships
To find novel applications for established therapeutics, we chose
to investigate if evidence could be accrued to indicate if
Tamoxifen, a selective estrogen receptor modulator approved for
breast cancer, might have additional uses. Some beneficial effects
of Tamoxifen on SLE (a chronic autoimmune disease that may
affect multiple organ systems) have been observed in animal
tests17 as well as some preliminary clinical studies, supporting the
hypothesis that selective estrogen receptor modulators such as
Tamoxifen may have therapeutic potential in SLE patient
management18, 19.
While the pathogenesis of SLE is complex and poorly
understood, we sought to identify connectivities offered via
representation of gene networks associated with perturbations of
its implicated cell, pathway, ontology and phenotype correlates
with those of Tamoxifen. First, we issued Oracle RDF queries to
retrieve Tamoxifen and SLE RDF subgraph respectively:
• For Tamoxifen, we developed RDF queries to ask the
complex question “retrieve all genes and their annotation
(interacting gene, pathway, and gene ontology) that associated
with Tamoxifen by acting as its drug target(s) or indication(s)”
• Similarly, for SLE, we developed RDF queries to “retrieve
disease genes, or genes interacting with or sharing pathways
with SLE disease gene as well as their annotation”
Each query returns a set of variable bindings matching to the
query parameters and each unique result produces a graph formed
from the triples matching the criteria. The components of the
resultant RDF subgraph were summarized in Table 1. As
expected, since the connection between “Tamoxifen” and “SLE”
is non-trivial, no association was detected in each individual RDF
subgraph. However, by combining the extracted subgraphs and
applying inference rules using GO and Disease subsumption
relationships, we were able to extract the implicit connections
between the two entities of interest. For example, one of the
shortest associations extracted is via a common biological process
“apoptosis” (GO_0006915) that are both traversed by PDCD1
and CDH1, two genes that are found to be associated with known
3
Qu. AX. et al.
Tamoxifen indication and SLE. Fig.4 shows all the embedded
associations extracted from the combined SLE-Tamoxifen RDF
graph, consisting 45 entity nodes with the minimal geodesics of 6
traversing between Tamoxifen and SLE. In addition, we also
borrowed the centrality analysis algorithm and approach20to
compute the key biological entities for the extracted RDF graphs.
The RDF triples were used as input for generating nodes and
edges. As a result, two critical genes were identified with high
ranking scores: ESR1 (estrogen receptor 1), AR (androgen
receptor). Based on literature mining, both genes are found to be
differentially expressed in SLE patients with an indicated role in
SLE pathogenesis or patient management18, 21, 22.
Table 1: Statistics of RDF Graph associated with Tamoxifen, SLE, and
Combined
RDF Graph
SLE
Tamoxifen Combined
Entities
Associations
114
121
695
947
768
1050
Fig 4: Implicit associations between “Tamoxifen” and “SLE” (entities
pointed by yellow arrows) consisting of 45 vertices, with minimal
geodesics of 6
2. CONCLUSION AND FUTURE WORK
We have presented a novel OWL-formalized ontology
framework for use in biomedical and pharmacological domain
applications. Our work to implement an integrated pharmacomegenome-diseasome RDF network based on this framework
suggests that the DDCO is effective and robust in knowledge
acquisition, integration, and inconsistency resolution. The
application scenarios we presented in this paper illustrates that
the DDCO framework and its supported RDF graph data model,
in combination with graph traversal and mining methods, can be
used in an exploratory context to formulate either initiating or
validating hypotheses. The scenarios can also be generalized to
other research questions in drug development area (see our prior
work23, 24) to support identifying new target or therapeutics. Our
current and planned work seeks to deepen knowledge capture
and mechanism modeling to further refine the reasoning
capability of the OWL/RDF model and include additional
dimensions such as genetic polymorphisms, mutations, deeper
4
clinical features, and diverse pharmacological properties and
principles of drug action. Doing so should greatly extend
sensitivity and specificity for individual patients. We also plan
to continuously evaluate and improve the framework in
conjunction with future expansion of the semantic infrastructure
by enabling expert review for a specific disease to model its
mechanisms and variations using entities and relations from the
DDCO.
REFERENCES
1. Ashburn TT, Thor KB. Drug repositioning: Identifying and developing new
uses for existing drugs. Nat Rev Drug Discov. 2004;3:673-683.
2. Hewett M, Oliver DE, Rubin DL, et al. PharmGKB: The pharmacogenetics
knowledge base. Nucleic Acids Res. 2002;30:163-165.
3. Kanehisa M, Goto S. KEGG: Kyoto encyclopedia of genes and genomes.
Nucleic Acids Res. 2000;28:27-30.
4. Wishart DS, Guo AC, et al. DrugBank: A comprehensive resource for in silico
drug discovery and exploration. Nucleic Acids Res. 2006;34:D668-72.
5. J. H. Tim Berners-Lee The semantic web. Sci Ameri Mag.;(284): p. 29-37.
6. Bodenreider O. The unified medical language system (UMLS): Integrating
biomedical terminology. Nucleic Acids Res. 2004;32:D267-70.
7. Kashyap V. The UMLS semantic network and the semantic web. AMIA Annu
Symp Proc. 2003:351-355.
8. de Coronado S, Haber MW, Sioutos N, Tuttle MS, Wright LW. NCI
thesaurus: Using science-based terminology to integrate cancer research results.
Medinfo. 2004;11:33-37.
9. ATC::http://www.whocc.no/atcddd/.
10. CACTE: http://ctep.cancer.gov/reporting/ctc.html.
11. Ashburner M, Ball CA, Blake JA, et al. Gene ontology: Tool for the
unification of biology. the gene ontology consortium. Nat Genet. 2000;25:25-29.
12. Stearns MQ, Wang AY.et al SNOMED clinical terms: Overview of the
development process and project status. Proc AMIA Symp. 2001:662-666.
13. Noy NF, Crubezy M, Fergerson RW, et al. Protege-2000: An open-source
ontology-development and knowledge-acquisition environment. AMIA Annu
Symp Proc. 2003:953.
14. Haarslev V, Möller R. Racer: An OWL reasoning agent for the semantic
web. Proceedings of the International Workshop on Applications, Products, and
Sevices of Web-based Support Systems. 2003;18:91.
15. Hamosh A, Scott AF, Amberger JS, Bocchini CA, McKusick VA. Online
mendelian inheritance in man (OMIM), a knowledgebase of human genes and
genetic disorders. Nucleic Acids Res. 2005;33:D514-7.
16. ORACLE: http://www.oracle.com/technology/tech/semantic_technologies.
17. Sthoeger ZM, Zinger H, Mozes E. Beneficial effects of the anti-oestrogen
tamoxifen on systemic lupus erythematosus of (NZBxNZW)F1 female mice are
associated with specific reduction of IgG3 autoantibodies. Ann Rheum Dis.
2003;62:341-346.
18. Cohen-Solal Diamond B. et al Sex hormones and SLE: Influencing the fate
of autoreactive B cells. Curr Top Microbiol Immunol. 2006;305:67-88.
19. Peeva E, Venkatesh J, Michael D, Diamond B. Prolactin as a modulator of B
cell function: Implications for SLE. Biomed Pharmacother. 2004;58:310-319.
20. Koschützki D, Richter S, et al Tenfelde-Podehl D, Zlotowski O, ed.
Centrality Indices. In Network Analysis: MethodologicalFoundations. Springer;
2005; No. Volume 3418 of LNCS Tutorial.
21. Strand V. Approaches to the management of systemic lupus erythematosus.
Curr Opin Rheumatol. 1997;9:410-420.
22. Inui A, Ogasawara H, Naito T, et al. Estrogen receptor expression by
peripheral blood mononuclear cells of patients with systemic lupus
erythematosus. Clin Rheumatol. 2007;26:1675-1678.
23. Gudivada R, Qu AX, Jegga AG, Neumann EK, Aronow BJ. A genome –
phenome integrated approach for mining disease genes using semantic web.
WWW2007Workshop on Healthcare and Life Sciences. 2007.
24. Qu AX, Aronow B. et al Semantic web-based data representation and
reasoning applied to disease mechanism and pharmacology. Proceedings of 2007
IEEE Bioinformatics and Biomedicine. 2007:131-143.
Vol. 00 no. 00 2008
Pages 1–4
BIOINFORMATICS
BOWiki: An ontology-based wiki for annotation of data and
integration of knowledge in biology
Robert Hoehndorf,a,c,d Joshua Bacher,b,c Michael Backhaus,a,c,d Sergio E. Gregorio,
Jr.,e Frank Loebe,a,d Kay Prüfer,c Alexandr Uciteli,c Johann Visagie,c Heinrich Herre
a,d
and Janet Kelso c
a Department
of Computer Science, Faculty of Mathematics and Computer Science, University of Leipzig,
Johannisgasse 26, 04103 Leipzig, Germany; b Institute for Logics and Philosophy of Science, Faculty of Social
Science and Philosophy, University of Leipzig, Beethovenstrasse 15, 04107 Leipzig, Germany; c Department of
Evolutionary Genetics, Max Planck Institute for Evolutionary Anthropology, Deutscher Platz 6, 04103 Leipzig,
Germany; d Research Group Ontologies in Medicine (Onto-Med), Institute of Medical Informatics, Statistics
and Epidemiology (IMISE), University of Leipzig, Härtelstrasse 16–18, 04107 Leipzig, Germany;
e Communications and Publications Services, International Rice Research Institute, College, 4030 Los Baños,
Laguna, Philippines
ABSTRACT
Ontology development and the annotation of biological data using
ontologies are time-consuming exercises that currently requires input from
expert curators. Open, collaborative platforms for biological data annotation
enable the wider scientific community to become involved in developing
and maintaining such resources. However, this openness raises concerns
regarding the quality and correctness of the information added to these
knowledge bases. The combination of a collaborative web-based platform
with logic-based approaches and Semantic Web technology can be used to
address some of these challenges and concerns.
We have developed the BOWiki, a web-based system that includes a
biological core ontology. The core ontology provides background knowledge
about biological types and relations. Against this background, an automated
reasoner assesses the consistency of new information added to the
knowledge base. The system provides a platform for research communities
such as wikis for the description, discussion and annotation of the
functions of genes and gene products [Wang, 2006, Hoehndorf et al.,
2006, Giles, 2007].
However, an open approach like wikis frequently raises concerns
regarding the quality of the information captured. The information
represented in the wiki should adhere to particular quality
criteria such as internal consistency (the wiki content does not
contain contradictory information) and consistency with biological
background knowledge (the wiki content should be semantically
correct). To address some of these concerns, logic-based tools can
be employed.
We have developed the BOWiki, a wiki system that uses a
core ontology together with an automated reasoner to maintain a
consistent knowledge base. It is specifically targeted at small- to
medium-sized communities.
to collaboratively integrate information and annotate data.
The BOWiki and supplementary material is available at http://www.
bowiki.net/. The source code is available under the GNU GPL from
http://onto.eva.mpg.de/trac/BoWiki.
Contact: [email protected]
1
INTRODUCTION
Biological ontologies have been developed for a number of
domains, including cell structure, organisms, biological sequences,
biological processes, functions and relationships. These ontologies
are increasingly being applied to describe biological knowledge.
Annotating biological data with ontological categories provides an
explicit description of specific features of the data, which enables
users to integrate, query and reuse the data in ways previously not
possible, thereby significantly increasing the data’s value.
Developing and maintaining these ontologies requires manual
creation, deletion and correction of concepts and their definitions
within the ontology, as well as annotating biological data to concepts
from the ontology. In order to overcome the arising acquisition
bottleneck, several authors suggest using community-based tools
c Oxford University Press 2008.
2
SYSTEM DESCRIPTION
The BOWiki is a semantic wiki based on the MediaWiki1 software.
In addition to the text-centered collaborative environment common
to all wikis, a semantic wiki provides the user with an interface for
entering structured data [Krötzsch et al., 2007]. This structured data
can be used subsequently to query the data collection. For instance,
inline queries [Krötzsch et al., 2007] can be added to the source
code of a wikipage, which will always produce an up-to-date list of
results on a wikipage.
The BOWiki significantly extends the MediaWiki’s capabilities.
It allows users to characterize the entities specified by wikipages as
instances of ontological categories, to define new relations within
the wiki, to interrelate wikipages, and to query for wikipages
satisfying some criteria. In particular, the BOWiki provides features
beyond those offered by common wiki systems (for details see
the Implementation section and table 1): typing wikipages (table
1), n-ary semantic relations among wikipages (table 1), semantic
1
http://www.mediawiki.org
1
Hoehndorf et al.
search (special page, inline queries), reasoner support for content
verification, adaptability to an application domain, import of
bio-ontologies for local accessibility and simple reuse, graphical
ontology browsing and OWL [McGuinness and van Harmelen,
2004] export of the wiki content.
We consider both adaptability to the application domain and
content verification as the BOWiki’s two most outstanding novel
features. Adaptability means that during setup, the software reads an
OWL ontology selected by the user that provides a type system for
the wikipages and the relations that are available to connect them.
New relations can be introduced using specific wiki syntax, while
the types remain fixed after setup.
While semantic wikis allow for the structured representation of
information, they often provide little or no quality control, and
do not verify the consistency of captured knowledge. Using the
imported ontology as a type system in the BOWiki enforces the use
of a common conceptualization and provides additional background
knowledge about the selected domain. This background knowledge
is used to check user-entered, semantic content by means of an
OWL reasoner. For example, the ontology can prevent typing an
instance of p45 both with Protein and DNA molecule at the same
time. Currently, the performance of automated reasoners remains a
limiting factor. Nevertheless, the reasoner delivers a form of quality
control for the BOWiki content that should be adopted wherever
possible.
The BOWiki was primarily designed to describe biological data
using ontologies. In conjunction with a biological core ontology
[Valente and Breuker, 1996] like GFO-Bio [Hoehndorf et al., 2007]
or BioTop [Schulz et al., 2006], the BOWiki can be used to describe
biological data. For this purpose, we developed a module that
allows OBO flatfiles2 to be imported into the BOWiki. By default,
these ontologies are only accessible for reading; they are neither
editable nor considered in the BOWiki’s reasoning. Users can then
create wikipages containing information about biological entities,
and describe the entities both in natural language text and in a
formally structured way. For the latter, they can relate the described
entities to categories from the OBO ontologies, and these categories
are then made available for use by the BOWiki reasoning.
In contrast to annotating data with ontological categories, i.e.,
asserting an undefined association relation between a biological
datum and an ontological category, it is possible in the BOWiki
to define precisely the relation between a biological entity (e.g. a
class of proteins) and another category: a protein may not only be
annotated to transcription factor activity, nucleus, sugar transport
and glucose. In the BOWiki, it may stand in the has function relation
to transcription factor activity; it can be located at a nucleus; it
can participate in a sugar transport process; it can bind glucose.
The ability to make these relations explicit renders annotations in a
semantic wiki both exceptionally powerful and precise.
The BOWiki can be used to describe not only data, but also
biological categories, or to create relations between biological
categories. As such, the BOWiki could be used to create so-called
cross-products [Smith et al., 2007] between different ontologies.
3
4
2
http://www.cs.man.ac.uk/˜horrocks/obo/
DISCUSSION
Using different reasoners
The BOWikiServer provides a layer of abstraction between the
description logic reasoner and the BOWiki. Depending on the
description logic reasoner used, different features can be supported.
Currently, the BOWikiServer uses the Pellet reasoner [Sirin and
Parsia, 2004]. Pellet supports the explanation of inconsistencies,
3
2
IMPLEMENTATION
Within our MediaWiki extension, users can specify the type of
entity described by a wikipage (see table 1). One of the central
ideas of the BOWiki is to provide a pre-defined set of types and
relations (and corresponding restrictions among them). We deliver
the BOWiki with the biological core ontology GFO-Bio, but any
consistent OWL [McGuinness and van Harmelen, 2004] file can be
imported as the type system. Types are modelled as OWL classes
and binary relations as OWL properties. Relations of higher arity
are modeled according to use case 3 in [Noy and Rector, 2006], i.e.,
as classes whose individuals model relation instances. Wikipages as
(descriptions of) instances of types give rise to OWL individuals,
which may be members of OWL classes (their types).
An OWL ontology can provide background knowledge about
a domain in the form of axioms that restrict the basic types and
relations within the domain. This allows for automatic verification
of parts of the semantic content created in the BOWiki: users
may introduce a new page in the wiki and describe some entity;
they may then add type information about the described entity;
and this added type information is then automatically verified.
The verification checks the logical consistency of the BOWiki’s
content – as OWL individuals and relations among them – with
the restrictions of the OWL ontology’s types and relations, like
those in GFO-Bio. The BOWiki uses a description logic [Baader
et al., 2003] reasoner to perform these consistency checks. We
implemented the BOWikiServer, a stand-alone server that provides
access to a description logic reasoner using the Jena 2 Semantic Web
Framework [Carroll et al., 2003] and a custom-developed protocol.
A schema of the BOWiki’s architecture is illustrated in figure 1.
Whenever a user edits a wikipage in the BOWiki, the consistency
of the changes with respect to the core ontology is verified using the
BOWikiServer. Only consistent changes are permitted. In the event
of an inconsistency, an explanation for the inconsistency is given,
and no change is made until the user resolves the inconsistency.
In addition to verifying the consistency of newly added
knowledge, the BOWikiServer can perform complex queries over
the data contained within the wiki. Queries are performed as
retrieval operations for description logic concepts [Baader et al.,
2003], i.e., as queries for all individuals that satisfy a description
logic concept description.
A performance evaluation of our implementation using the Pellet
description logic reasoner [Sirin and Parsia, 2004] for ontology
classification showed, that presently, only small- to medium-sized
wiki installations can be supported. The time needed for consistency
checks increases as the number of wiki pages increases3 .
The results of our performance tests can be found on the wiki at http:
//bowiki.net.
BOWiki
BOWiki syntax
Generic
1 [[OType:C]]
2 [[R::page2]]
3 [[R::role1=page1;...;roleN=pageN]]
4
[[has-argument::
name=roleName;type=OType:C]]
Examples
1 on page Apoptosis: [[OType:Category]]
2 on page Apoptosis:
[[CC-isa::Biological process]]
3
on page HvSUT2:
[[Realizes::
function=Sugar transporter activity;
process=Glucose transport]]
4
OWL abstract syntax
Individual(page type(C))
Individual(page value(R page2))
Individual(R-id type(R))
Individual(R-id value(subject page))
Individual(R-id value(R-role1 page1))
...
Individual(R-id value(R-roleN pageN))
SubClassOf(page gfo:Relator)
ObjectProperty(R-roleName domain(page) range(C))
Individual(Apoptosis, type(Category))
Individual(Apoptosis value(CC-isa Biological process))
Individual(Realizes-0 type(Realizes))
Individual(Realizes-0 value(Realizes-subject HvSUT2))
Individual(Realizes-0 value(Realizes-function Sugar transporter activity))
Individual(Realizes-0 value(Realizes-process Glucose transport))
on page Realizes:
[[has-argument::
name=function;
type=OType:Function category]]
SubClassOf(Realizes gfo:Relator))
ObjectProperty(Realizes-function domain(Function category))
Table 1. Syntax and semantics of the BOWiki extensions. The table shows the syntax constructs used in the BOWiki for semantic markup. The second column
provides a translation into OWL. (page refers to the wikipage in which the statement appears; “R-id” is a name for an individual whose “id” part is unique
and generated automatically for the occurrence of the statement). Because OWL has a model-theoretic semantics, this translation yields a semantics for the
BOWiki syntax. In the lower half of the table we illustrate each construct with an example and present its particular translation to OWL.
which can be shown to users to help them in correcting
inconsistent statements submitted to the BOWiki. It also supports
the nonmonotonic description logic ALCK with the auto-epistemic
K operator [Donini et al., 1997]. This permits both open- and
closed-world reasoning [Reiter, 1980] to be combined, which
has several practical applications in the Semantic Web [Grimm
and Motik, 2005] and the integration of ontologies in biology
[Hoehndorf et al., 2007]. On the other hand, reasoning in the OWL
description logic fragment [McGuinness and van Harmelen, 2004]
is highly complex. It is possible to use reasoners for weaker logics
to overcome the performance limitations encountered with Pellet.
Comparison with other approaches
WikiProteins [Giles, 2007] is a software project based also on the
MediaWiki software, focused on annotating Swissprot [Boeckmann
et al., 2003]. Similar to the BOWiki, it utilizes ontologies like the
Gene Ontology [Ashburner et al., 2000] and the Unified Medical
Language System [Humphreys et al., 1998] as a foundation for the
annotation. It is generally more targeted at creating and collecting
definitions for terms than on formalizing knowledge in a logic-based
and ontologically founded framework. As a result, it contains a
mashup of lexical, terminological and ontological information. In
addition, WikiProteins neither supports n-ary relations nor provides
a description logic reasoner to retrieve or verify information. It
therefore lacks the quality control and retrieval features that are
central to the BOWiki. On the other hand, because of the different
use-cases that WikiProteins supports, it is designed to handle much
larger quantities of data than the BOWiki, and it is better suited for
creating and managing terminological data.
The Semantic Mediawiki [Krötzsch et al., 2007] is another
semantic wiki based on the Mediawiki software. It is designed to
be applicable within the online encyclopedia Wikipedia. Because of
the large number of Wikipedia users, performance and scalability
requirements are much more important for the Semantic Mediawiki
than for the BOWiki. Therefore, it also provides neither a
description logic reasoner nor ontologies for content verification.
The IkeWiki [Schaffert et al., 2006], like the BOWiki,
includes the Pellet description logic reasoner for classification and
verification of consistency. In contrast to the BOWiki, parts of the
IkeWiki’s functionality require users to be experts in either Semantic
Web technology or knowledge engineering. As a consequence, the
BOWiki lacks some of the functionality that the IkeWiki provides
(such as creating and modifying OWL classes) as it targets biologist
users, most of whom are not trained in knowledge engineering.
Conclusion
We developed the BOWiki as a semantic wiki specifically designed
to capture knowledge within the biological and medical domains.
It has several features that distinguish it from other semantic wikis
and from similarly targeted projects in biomedicine, most notably
its ability to verify its semantic content for consistency with respect
to background knowledge and its ability to access external OBO
ontologies.
3
Hoehndorf et al.
Semantic
components
Internet
Web server (Apache) + PHP environment
Presentation
components
Components
developed as
part of this
project
BOWiki
extension
TCP/IP socket
MediaWiki
(a)
Relational database (MySQL)
BoWikiServer
Jena 2 Semantic Web
framework
Description logic reasoner
(Pellet)
MediaWiki
database
BOWiki database
extensions
(c)
(b)
import
OWL-DL ontology
Fig. 1: BOWiki Architecture. (a) The BOWiki extension to
the MediaWiki software processes the semantic data added to
wiki pages. The semantic data is subsequently transferred to
the BOWikiServer using a TCP/IP connection. (b) To evaluate
newly entered data or semantic queries, the BOWikiServer
requires an ontology in OWL-DL format (provided during
installation of the BOWiki). Consistent semantic data will
be stored. If an inconsistency is detected, the edited page
is rejected with an explanation of the inconsistency. The
BOWikiServer currently uses the Jena 2 Semantic Web
framework together with the Pellet reasoner. (c) After
successful verification the semantic data is stored in a separate
part of the SQL database.
The BOWiki allows a scientific community to annotate biological
data rapidly. This annotation can be performed using biomedical
ontologies. In addition to data annotation, the specific type of
relations between entities can be made explicit. It is also possible
to integrate different biological knowledge bases by creating partial
definitions for the relations and categories used in the knowledge
bases.
The BOWiki employs a type system to verify the consistency of
the knowledge represented in the wiki. The type system is provided
in the form of an OWL knowledge base. If the type system is a
core ontology for a domain (i.e., it provides background knowledge
and restrictions about the categories and relations for the domain),
its use contributes to maintaining the ontological adequacy of the
BOWiki’s content, and thereby the content’s quality.
4
ACKNOWLEDGEMENTS
We thank Christine Green for her help in preparing the English
manuscript.
REFERENCES
M. Ashburner, C. A. Ball, J. A. Blake, D. Botstein, H. Butler, J. M. Cherry, A. P.
Davis, K. Dolinski, S. S. Dwight, J. T. Eppig, and et al. Gene ontology: tool for the
unification of biology. The Gene Ontology Consortium. Nat Genet, 25(1):25–29,
2000.
F. Baader, D. Calvanese, D. McGuinness, D. Nardi, and P. Patel-Schneider, editors.
The Description Logic Handbook: Theory, Implementation and Applications.
Cambridge University Press, Cambridge, UK, 2003.
B. Boeckmann, A. Bairoch, R. Apweiler, M.-C. Blatter, A. Estreicher, E. Gasteiger,
M. J. Martin, K. Michoud, C. O’Donovan, I. Phan, and et al.. The SWISS-PROT
protein knowledgebase and its supplement TrEMBL in 2003. Nucleic Acids Res, 31
(1):365–370, January 2003.
J. J. Carroll, I. Dickinson, C. Dollin, D. Reynolds, A. Seaborne, and K. Wilkinson.
Jena: Implementing the Semantic Web recommendations. Technical Report HPL2003-146, Hewlett Packard, Bristol, UK, 2003.
F. M. Donini, D. Nardi, and R. Rosati. Autoepistemic description logics. In M. E.
Pollack, editor, Proceedings of the Fifteenth International Joint Conference on
Artificial Intelligence, IJCAI 1997, Nagoya, Japan, Aug 23-29, volume 1, pages
136–141, San Francisco, 1997. Morgan Kaufmann.
J. Giles. Key biology databases go wiki. Nature, 445(7129):691, 2007.
S. Grimm and B. Motik. Closed world reasoning in the Semantic Web through
epistemic operators. In B. Cuenca Grau, I. Horrocks, B. Parsia, and P. PatelSchneider, editors, Proceedings of the OWLED’05 Workshop on OWL: Experiences
and Directions, Galway, Ireland, Nov 11-12, volume 188 of CEUR Workshop
Proceedings, Aachen, Germany, 2005. CEUR-WS.org.
R. Hoehndorf, K. Prüfer, M. Backhaus, H. Herre, J. Kelso, F. Loebe, and J. Visagie. A
proposal for a gene functions wiki. In R. Meersman, Z. Tari, and P. Herrero, editors,
Proceedings of OTM 2006 Workshops, Montpellier, France, Oct 29 - Nov 3, Part
I, Workshop Knowledge Systems in Bioinformatics, KSinBIT 2006, volume 4277 of
Lecture Notes in Computer Science, pages 669–678, Berlin, 2006. Springer.
R. Hoehndorf, F. Loebe, J. Kelso, and H. Herre. Representing default knowledge in
biomedical ontologies: Application to the integration of anatomy and phenotype
ontologies. BMC Bioinformatics, 8(1):377, 2007.
B. L. Humphreys, D. A. B. Lindberg, H. M. Schoolman, and G. O. Barnett. The Unified
Medical Language System: an informatics research collaboration. J Am Med Inform
Assoc, 5(1):1–11, 1998.
M. Krötzsch, D. Vrandečić, M. Völkel, H. Haller, and R. Studer. Semantic wikipedia.
Web Semantics: Science, Services and Agents on the World Wide Web, 5(4):251–261,
2007.
D. L. McGuinness and F. van Harmelen. OWL Web Ontology Language overview.
W3C recommendation, World Wide Web Consortium (W3C), 2004.
N. Noy and A. Rector. Defining N-ary relations on the Semantic Web. W3C working
group note, World Wide Web Consortium (W3C), 2006.
R. Reiter. A logic for default reasoning. Artificial Intelligence, 13(1-2):81–132, 1980.
S. Schaffert, R. Westenthaler, and A. Gruber. IkeWiki: A user-friendly semantic wiki. In
H. Wache, editor, Demos and Posters of the 3rd European Semantic Web Conference,
ESWC 2006, Budva, Montenegro, Jun 11-14, 2006.
S. Schulz, E. Beisswanger, J. Wermter, and U. Hahn. Towards an upper-level ontology
for molecular biology. AMIA Annu Symp Proc, 2006:694–698, 2006.
E. Sirin and B. Parsia. Pellet: An OWL DL reasoner. In V. Haarslev and R. Möller,
editors, Proceedings of the 2004 International Workshop on Description Logics,
DL2004, Whistler, British Columbia, Canada, Jun 6-8, volume 104 of CEUR
Workshop Proceedings, pages 212–213, Aachen, Germany, 2004. CEUR-WS.org.
B. Smith, M. Ashburner, C. Rosse, J. Bard, W. Bug, W. Ceusters, L. J. Goldberg,
K. Eilbeck, A. Ireland, C. J. Mungall, N. Leontis, P. Rocca-Serra, A. Ruttenberg, S.A. Sansone, R. H. Scheuermann, N. Shah, P. L. Whetzel, and S. Lewis. The OBO
Foundry: coordinated evolution of ontologies to support biomedical data integration.
Nat Biotech, 25(11):1251–1255, 2007.
A. Valente and J. Breuker. Towards principled core ontologies. In B. R. Gaines and
M. A. Musen, editors, Proceedings of the 10th Knowledge Acquisition Workshop,
KAW’96, Banff, Alberta, Canada, Nov 9-14, pages 301–320, 1996.
K. Wang. Gene-function wiki would let biologists pool worldwide resources. Nature,
439(7076):534, 2006.
PubOnto: Open Biomedical Ontology-Based Medline Exploration
Weijian Xuan, Manhong Dai, Brian Athey, Stanley J. Watson and Fan Meng*
Psychiatry Department and Molecular and Behavioral Neuroscience Institute, University of Michigan, US
ABSTRACT
Motivation: Effective Medline database exploration is critical for the understanding of high throughput experimental
results and the development of biologically relevant hypotheses. While existing solutions enhance Medline exploration
through different approaches such as document clustering,
network presentation of underlying conceptual relationships
and the mapping of search results to MeSH and Gene Ontology trees, we believe the use of multiple ontologies from
the Open Biomedical Ontology can greatly help researchers
to explore literature from different perspectives as well as
quickly locate the most relevant Medline records.
Availability: The PubOnto prototype is freely accessible at:
http://brainarray.mbni.med.umich.edu/brainarray/prototype/p
ubonto
1
INTRODUCTION
The popularity of data driven biomedical research leads
to large volumes of data such as gene expression profiles,
MRI images and SNPs related to various pathophysiological
processes. As a result, understanding the biological implications of the high throughput data has become a major challenge (Boguski and McIntosh, 2003). It requires timeconsuming literature and database mining and is the main
goal of the “literature-based discovery”, “conceptual biology”, or more broadly, “electronic biology”, through which
biologically important hypotheses are derived from existing
literature and data using various approaches (Jensen, et al.,
2006; Srinivasan, 2004; Swanson, 1990; Wren, et al., 2004).
The effectiveness of such knowledge mining also relies
heavily on researchers’ background knowledge about novel
genes or SNPs, and this knowledge, at present, is sparse.
The Medline database is without doubt the foremost
biomedical knowledge database that plays a critical role in
the understanding of high throughput data. Unfortunately,
prevailing Medline search engines such as PubMed and
Google Scholar have been designed largely for the efficient
retrieval of a small number of records rather than an indepth exploration of a large body of literature for discovery
and proof purposes. They rely heavily on a step-wise narrowing of search scope but such an approach does not work
well for the exploration of uncharted territories. This is because background knowledge is needed for defining sensible
*
To whom correspondence should be addressed.
filtering criteria and guessing what are potentially relevant
topics for additional exploration. For example, in microarray
gene expression analysis, researchers frequently have to
deal with lists of genes that are not known to be associated
with the targeted biological processes. Researchers have to
utilize other intermediate concepts to establish indirect links
between gene lists and specific biological processes. However, identifying such intermediate concepts is very difficult
in existing solutions and it is not easy even in systems devoted for this purpose such as ArrowSmith (Smalheiser, et
al., 2007; Swanson, 1986). Frequently researchers have to
go through large number of retrieved records one-by-one
and examine external databases to find interesting new relationships.
Another major shortcoming for prevailing search solutions is that they do not present results in the contexts that a
user maybe interested in. For example, Google Scholar/PubMed basically present search results as a linear list of
papers. Users do not know the context of each paper nor the
relationship among these papers. Besides the original ranking provided by the search engine, there is little additional
cues and sorting/filtering methods that can facilitate the exploration of search results.
We believe the projection of search results to existing
knowledge structure is very important for hypothesis development. This is because researchers often need to explore
unfamiliar fields in the age of high throughput experiments
and such projections can provide much needed guidance in
new areas. In fact, even if a researcher wants to examine
related facts in his/her own field, there are many details related to the search topic that require additional efforts to
retrieve. Mapping search results to knowledge structures
will also be very useful for revealing hidden relationships
not easily identified by prevailing approaches. For example,
if Medline search results show several genes in a brain region are related to a disease in a statistically significant
manner, it will be worthwhile to explore the relationship of
other genes expressed in this brain region with the disease.
Naturally, exploration of multiple knowledge structures is
often needed to facilitate the formation of new insights. The
projection of search results to multiple dynamically-linked
knowledge structures is thus necessary for such contextassisted data and literature exploration.
Newer Medline search solutions such as GoPubMed
(Doms and Schroeder, 2005) and Vivisimo (Taylor, 2007)
1
Xuan et al.
attempts to organize search results in the context of either
predefined ontology such as Gene Ontology or dynamically
generated ontology structures based on clustering results. In
such solutions, users can rely on the tree-like organization
of search results to easily navigate to topics of interest. The
neighborhood of a given tree branch automatically suggests
related topics for additional exploration. Here predefined
ontologies have an advantage over clustering results for
exploring unfamiliar territories due to their systematic listing of related concepts and their relationships.
However, given the huge number of biomedical concepts (e.g., over 1 million in the Unified Medical Language
System) and the complexity of relationships among them, it
is not possible to rely on one or two ontologies for effective
exploration. Researchers must have the capability to examine their search results from different perspectives. We
believe an ontology-based Medline exploration solution
must allow the use of different orthogonal ontologies, i.e.,
ontologies that addressing different aspects of biomedical
research. In addition, it is critical to enable interactive filtering of search results using terms from different ontologies
for more efficient Medline exploration.
The main goal of this work is to develop a flexible ontology-based Medline exploration solution to facilitate the
understanding of high throughput data and the discovery of
potentially interesting conceptual relationships. Our solution
enables interactive exploration of search results through the
use of multiple ontologies from OBO foundry. It also has an
open architecture that allows flexible selection of Medline
retrieval algorithms through different web services.
2
User Search
Ontology selection
FMA
MeSH
e-utilities
GO
other services
Web service layer
Text mining
PubMed
Integration
Gene
ENVO
…
METHODS
Selection of Ontologies: The Open Biomedical Ontologies
(OBO) foundry is a comprehensive collaborative effort to
create controlled vocabularies for shared use across different
biological and medical domains (NCBO, 2008) (Rubin, et
al., 2006; Smith, et al., 2007). It already includes around 50
ontologies from various biomedical domains. We selected
Gene Ontology, Foundational Model of Anatomy, Mammalian Phenotype Ontology and Environment Ontology for
inclusion in our prototype since they provide key perspectives for topics of great interest for biomedical research and
they are almost orthogonal to each other conceptually.
Mapping of Ontology to Medline: We developed a very
efficient general purpose ontology to free-text mapping solution in collaboration with researchers in the National Center for Biomedical Ontology. In brief, our solution relies on
the pre-generation of lexical variations, word order permutations for ontology terms, their synonyms together with a
highly efficient implementation of a suffix-tree based string
match algorithm. Our solution is able to map all concepts in
UMLS to the full Medline database in 15 hours on a mainstream Opteron server. It achieves over 95% recall rate
2
when compared to the results from the MMTx program,
which is about 500 times slower and does not support the
use of non-UMLS ontologies. The details of our ontology
mapping solution will be presented in a separate paper.
PubOnto Architecture: In order to provide a web-based
Medline exploration tool with rich interactivity, we developed PubOnto on Adobe’s latest Flex 3.0 platform. It allows
us to build highly interactive user interface that is compatible in virtually all major browsers. We developed an innovative technique that dynamically updates the XML-based
ontology tree structure by building a web service for each
ontology for feeding expanded nodes with ontological information and literature searching results. As a result, only a
minimum amount of data is transferred asynchronously and
PubOnto can thus handle very large ontologies. Fig. 1
shows the architecture of PubOnto. Since the web service
layer separates the user interface from ontologies, search
services and other databases, the back end changes do not
affect the client side user interface.
Ontology exploration
FMA
GO
…
…
ENVO
Scaffolding tools:
Citation
MeSH analysis
Visualization
Fig. 1. PubOnto architecture
3
RESULTS
PubOnto is a FLEX application providing high level of
interactivity for efficient Medline search result exploration.
We illustrate a number of key features in this section.
Ontology-based exploration of search results: Simply
displaying search results for each individual node is often
not satisfactory. Typically users want to know quickly how
many literatures are retrieved for all children under a branch
so that they can decide if something is interesting that needs
to be explored further. Rolling up such mapping data in a
large ontology such as FMA or GO on-the-fly is not an easy
task. Traditional tree traversal algorithms are very CPU intensive and usually require large in-memory tree structures
PubOnto: Open Biomedical Ontology-Based Medline Exploration
on the server. To provide real time interactivity, we pretraverse the entire ontology and generate a parent-child table
that matches all nodes in the subtree to their parent nodes.
We also save the literature retrieval results to a sessionbased table. When a user expands a node, our service will
perform an efficient table join to obtain the aggregated information.
Ontology Selection: PubOnto support a series of OBO ontologies. However, we also understand that users may not
need to examine all of them. Therefore, we present a flexible way for users to choose which of the supported ontologies they want to use (Fig. 2). Once a user selects certain
ontologies, PubOnto will dynamically create a new ontology
tab with consistent display and interactive functions. We are
also developing functions that can use multiple ontologies
as combined filters to better navigate through citations.
Scaffolding Tools: PubOnto provides a number of tools for
easy exploration. When a user clicks an ontology node, corresponding citations will show up in the bottom panel.
Clicking on each citation will bring up a dialog for detailed
citation information (Fig. 4). PubOnto also provide aggregated MeSH information to highlight concepts that are significant in this particular citation set versus the whole Medline corpus. In addition, PubOnto provides charting function
to visualize MeSH concept distribution in search results.
Fig. 4. Citation exploration
Fig. 2. Ontology selection
Search Result Exploration: When a user submits a keyword search request, web services will return retrieval results for each selected ontology. The user can expand tree
nodes to explore results, as show in Fig. 3.
Fig. 3. Ontology-based exploration
While the most important feature of PubOnto is the
ability to use multiple OBO ontologies for Medline exploration, it also offers a number of unique features summarized
in Table 1. The PubOnto prototype currently does not include several functions in GoPubMed that are not directly
related to ontology but similar functions will be added in
future upon users’ request.
Table 1. Comparison between PubOnto and GoPubMed
PubOnto GoPubMed
More ontologies besides MeSH, GO
Interaction among ontology
Customizable search service
Client side filter
Customizable ontology search
Customizable interaction function
Rich interactions
Search history maintenance
Sorting citation by various criteria
Export to Citation managers
Citation linkouts
Where/Where/When analysis
Keyword highlight
Hot topics
Wikipedia mapping
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Direct
Yes
No
No
No
No
No
No
No
No
No
No
No
Yes
No
Indirect
Yes
Yes
Yes
Yes
Yes
3
Xuan et al.
4
DISCUSSIONS
Systematic ontology development efforts such as those
related to the Open Biomedical Ontologies are likely to generate expansive conceptual framework for the integration,
analysis and understanding of data generated in different
areas of biomedical research. PubOnto aims to capitalize on
the impressive progresses in ontology development for the
exploration and mining of biomedical literature. The ability
to utilize multiple orthogonal ontologies during Medline
exploration can significantly increase the efficiency of locating interesting search results in areas that researchers are not
familiar with. Mapping Medline results to multiple ontologies also enables researchers to explore search results from
different angles for new hypothesis development.
While the PubOnto prototype provides a conceptual
demo for the power of using multiple ontologies for Medline exploration, there are a number of improvements we
hope to incorporate in the coming months. For example,
although the ability to select different ontologies for organizing search results is quite powerful, it is based on the
assumption that users know which ontologies they want to
use. It should be possible to rank ontologies for their usefulness to the topic based on distribution of returned Medline
records on different concepts under a given ontology. For
example, an ontology is not very useful for Medline search
result exploration if only a small fraction of returned records
can be mapped to this ontology. On the contrary, an ontology will be very effective if many records can be mapped to
it and those records are relatively evenly distributed across
many terms in that ontology. Of course, an ontology is still
not useful if most of the search results can be mapped to
only a few terms in an ontology. Consequently, it should be
possible to develop an ontology scoring system based on the
number of records that can be mapped to an ontology and
the distribution of Medline records in an ontology for the
automatic selection of default ontology for a given Medline
search result. Conceivably, once the first ontology is selected, it is possible to select the second best ontology based
on the “orthogonality” with the first ontology. Of course,
such automated ontology ranking procedures are only based
on the statistical properties of the Medline records to ontology mapping. Users’ biomedical knowledge and their understanding of different ontologies will be essential for effective exploration of Medline literature.
Similarly, the exploration of a given ontology tree currently is also dependent on users background knowledge
since only the number of Medline records hits for a given
term can be used as external cues for ontology exploration
now. If there are many different ontologies for a user to
select from or the user is not familiar with the corresponding
ontology at all, it is desirable to have additional information
to help users to use such ontology guided exploration more
effectively. It is conceivable that we can weigh the specifici-
4
ty of each ontology term based on their inverse frequency of
showing up in Medline corpus so that users can focus on
more specific terms rather than exploring generic terms.
In summary, we believe the use of multiple ontologies
in OBO for Medline exploration can significantly increase
the efficiency of Medline exploration and facilitate the examination of the same search result from different perspectives. We will continue to improve PubOnto to make it an
effective tool for novel biomedical hypotheses development,
and ultimately incorporate it into PubViz, our more comprehensive biomedical literature exploration engine.
ACKNOWLEDGEMENTS
W. Xuan, M. Dai, S. J. Watson and F. Meng are members of
the Pritzker Neuropsychiatric Disorders Research Consortium, which is supported by the Pritzker Neuropsychiatric
Disorders Research Fund L.L.C. This work is also partly
supported by the National Center for Integrated Biomedical
Informatics through NIH grant 1U54DA021519-01A1 to the
University of Michigan
REFERENCES
Boguski, M.S. and McIntosh, M.W. (2003) Biomedical informatics
for proteomics, Nature, 422, 233-237.
Doms, A. and Schroeder, M. (2005) GoPubMed: exploring
PubMed with the Gene Ontology, Nucleic Acids Res, 33,
W783-786.
Jensen, L.J., Saric, J. and Bork, P. (2006) Literature mining for the
biologist: from information retrieval to biological discovery,
Nature Reviews Genetics, 7, 119-129.
NCBO
(2008)
The
Open
Biomedical
Ontologies,
http://www.obofoundry.org/.
Rubin, D.L., Lewis, S.E., , et al (2006) National Center for Biomedical Ontology: advancing biomedicine through structured
organization of scientific knowledge, Omics, 10, 185-198.
Smalheiser, N.R., Torvik, V.I., et al. (2007) Arrowsmith,
http://arrowsmith.psych.uic.edu/arrowsmith_uic/index.html.
Smith, B., Ashburner, M., Rosse, C., et al (2007) The OBO Foundry: coordinated evolution of ontologies to support biomedical
data integration, Nat Biotechnol, 25, 1251-1255.
Srinivasan, P. (2004) Text mining: Generating hypotheses from
MEDLINE, JASIST, 55, 396-413.
Swanson, D.R. (1986) Fish oil, Raynaud's syndrome, and undiscovered public knowledge, Perspect Biol Med, 30, 7-18.
Swanson, D.R. (1990) Medical literature as a potential source of
new knowledge, Bull Med Libr Assoc, 78, 29-37.
Taylor, D.P. (2007) An integrated biomedical knowledge extraction and analysis platform: using federated search and document clustering technology, Methods Mol Biol, 356, 293-300.
Wren, J.D., Bekeredjian, R., et al (2004) Knowledge discovery by
automated identification and ranking of implicit relationships,
Bioinformatics, 20, 389-398.
Minimal Anatomy Terminology (MAT): a species-independent
terminology for anatomical mapping and retrieval
Jonathan B.L. Bard1, James Malone2, Tim F. Rayner2 and Helen Parkinson2
1. Computational Biology Research Group, Weatherall Institute for Molecular Medicine, John Radcliffe Hospital, Oxford
OX3 9DS, UK
2. EMBL-EBI, Wellcome Trust Genome Campus, Hinxton, Cambridge, CB10 1SD, UK
ABSTRACT
Motivation: The problem of integrating a multiplicity of
non-orthogonal anatomy ontologies is well known in ontology development. There are now major public ontology
repositories (e.g. the Ontology for Biomedical Ontologies)
that require a multi-species anatomy ontology. We present
MAT (Minimal Anatomy Terminology) an OBO format terminology (~400 terms) using SKOS broader-than relationships designed for annotating and searching tissueassociated data and timelines for any organism. Identifiers
from >20 anatomy ontologies are mapped to each MAT
term to facilitate access to and interoperability across tissue-associated data resources
Availability: www.ebi.ac.uk/microarray-srv/mat/
1
INTRODUCTION
Data in public biomedical databases typically has various
classes of metadata has associated with it that enable
searching and analysis, and standards for different data
types and domains are now becoming available (e.g. a
series of Minimum Information protocols for this purpose,
mibbi.sourceforge.net/resources.shtml). There is no such
minimal standard for annotating anatomy because tissues
are much harder than other (e.g. experimental) data types
to formalize simply. This is partly because organisms
have so many diverse tissues and partly because tissue
organization is so complex. Nevertheless, because of the
need to handle tissue-associated data in databases, user
communities for all of the main model organisms have
produced formalized and fairly complete anatomical hierarchies (ontologies) that are largely based on part_of and
is_a relationships (Bard, 2005, 2007; Smith et al., 2007;
Burger et al., 2007). These high-granularity ontologies are
complex and their use presupposes considerable anatomical knowledge of the organism whose anatomy is
represented, as well as some understanding of the representation format of the ontology. They are therefore mainly used by specialist curators annotating data for the main
model organism databases using rich annotation tools
(e.g. Phenote, www.phenote.org).
Elsewhere, anatomical annotation is essentially free text,
or at best loosely controlled. Databases such as those from
the NCBI (www.ncbi.nlm.nih.gov) typically do not control anatomical annotation, and text mining is needed to
extract any anatomical information. It is unrealistic for
these multi-species databases archiving high throughput
data to develop annotation tools that provide intuitive
access to all (anatomy) ontologies and expect biomedical
users to use them consistently. ArrayExpress
(www.ebi.ac.uk/microarray-as/aer/entry), for example,
uses a text-mining strategy and string-matching methodology that adds no burden at the point of submission but
does require representative ontologies for automated annotation (Parkinson et al, 2006).
For query purposes, a simple anatomical ontology is
needed that allows searching and tree browsing, with its
complexity limited to that which is comprehensible to a
bench biologist. The simplest format for accessing annotation terms is a controlled vocabulary or terminology
where informal relationships connect the terms (unlike an
ontology whose formal relationships carry inheritance
implications). Two such terminologies are currently
available: the eVOC terminology set (Kelso et al., 2003)
whose scope is limited to human and mouse, and the very
short SAEL terminology (Parkinson et al., 2004) mainly
intended for core mammalian anatomical annotation and
which has no relations at all. Neither resource includes
identifiers for other anatomy-based resources that can be
used for cross-mapping and interoperability purposes.
This paper reports the development and validation of a
terminology entitled MAT (Minimal Anatomy Terminology). It is similar in format to eVOC but expanded to include high-level tissues and timelines appropriate for the
great majority of taxa rather than just mammals. Data
associated with these tissue terms include synonyms and
ontology identifiers for tissues from other anatomical ontologies currently downloadable from the Open Biomedical Ontologies (OBO) website (obofoundry.org/), and is
thus compatible with them.
1
J. Bard et al.
Figure 1 MAT terminology displayed in the CoBrA editor (www.xspan.org/cobra). The left panel shows the four top categories with
'anatomy basic component' expanded. The 'eye' has been expanded in the middle panel to show parent and child terms, synonyms
and identifiers. The right panel shows the organizing classes ‘taxon ontology’ and ‘time stages’
The MAT terminology is designed to facilitate the easy
annotation, curation and searching of tissue-associated
data while the ready availability of the various ontology
identifiers will facilitate tissue-associated interoperability
across databases
2
METHODS & RESULTS
Scope
The MAT terminology is intended to cover the basic
anatomy for all common taxa from fungi to plants and
animals to support anatomical information and mapping
of data contained in existing public resources. MAT is not
intended to represent formal knowledge about all these
organisms with its inherent implications for inheritance.
Identifiers
Each term has an identifier of the form MAT:0000001,
and is mapped to one or more identifiers from the anatomy ontologies currently available in the OBO foundry
(Smith et al., 2007). It also contains identifiers from the
Mammalian Phenotype Ontology (Smith, 2004) as these
typically include anatomical information and may be useful in the context of mapping abnormal phenotypes within
an anatomical context.
2
Granularity
Determining the appropriate level of granularity for MAT
is critical: too light and its archiving and searching uses
would be inadequate; too heavy a complexity would prohibit use by non-anatomists. The main indicator for tissue
selection is that formal species-specific ontologies (Drosophila, mouse etc) include these terms at a high level in
their respective representations. A second indicator is that
the selected tissues should be accessible for molecular
analysis. A third was that their meaning was obvious and
unambiguous to a biological user.
The current version of MAT has ~400 anatomical child
terms of the class anatomy basic component (Fig. 1). The
majority of these are used in their stage-independent
form. This is possible as most of the external ontologies
to which MAT is mapped have either restricted their
scope to adults or are structured so time and tissue are
handled independently (Burger et al., 2003).
Organizing principles
The terminology is intended to be intuitively navigable by
a biologist, and obvious choices for high level terms in
the hierarchy were organ and major tissue systems as
these both underpin anatomical organization in an intui-
Minimal Anatomy Terminology (MAT): a species-independent terminology for anatomical mapping and retrieval
tive way and are used by most anatomical ontologies.
MAT includes ~300 animal, ~75 pant and ~20 fungal
systems and tissues, Where tissues naturally fall into more
than one system (e.g. the mouth is both a craniofacial
tissue and part of the alimentary system), multiple inheritance has been used.
MAT includes two high level nodes in addition to anatomy basic component: taxon ontology, and time stage (Fig.
1). Detailed staging for each organism is outside the scope
of MAT, but a generic set of 11 stages each for animals
and plants that extend from the zygote to adult are used.
This allows distinctions to be made between, for example,
the embryonic, the juvenile and the adult testes.
It should be noted that a few ontologies (e.g. adult human
and adult mouse) only handle adult tissues, and their identifiers should not be used for developmental tissues.
As the MAT terminology is designed for mapping and
annotation rather than for logical inference, the use of
formal relationships such as is-a and part-of were replaced by the single broader-than relationship used as
defined by SKOS, the Simple Knowledge Organization
System, (www.w3.org/2004/02/skos/, a part of the Semantic Web (www.w3.org/2001/sw/). This allows us to
represent the terminology as a tree with a single, informal
relationship carrying no inheritance implications.
The MAT terms are intended to be species-independent,
trachea in the respiratory system has the associated identifiers from the Drosophila, human and mouse anatomy
ontologies even though the insect and vertebrate tracheae
are very different – they are analogues and not homologues. A sensu tag is used in only twice: the vertebrate and
invertebrate limbs are so different in structure and development that it seemed unreasonable to include them under
the same term, while the insect and amphibian fat bodies
are neither homologues nor analogues. MAT also contains
some transitional development-specific tissues with no
timing details (e.g. somite). These terms were included as
they would thus not be present in adult organism lists.
Different anatomy ontologies use different terms and
spellings for what are essentially equivalent terms (e.g.
oesophagus and esophagus, digestive system and alimentary system). We have made a subjective decision to use
the most common term as the standard (e.g. eye rather
than visual system, see Fig. 1), but synonyms are included
in the file and can be searched. In assigning identifiers
from other anatomy ontology to MAT terms (~1600 in
all), there was sometimes a choice as to which term to
map to. In the Drosophila ontology, for example, there is
a term for the digestive system and sub-terms for the embryonic/larval digestive systems and the pupal/adult digestive systems. Where alternatives exist the broadest
term is used preferentially. A very few tissues have been
included that are not present in other anatomical ontologies as they may be interesting in a wider evolutionary
context (e.g. phyllid, the gametophyte leaf).
The MAT terminology has very few text definitions as
almost all the terms are in common use by biologists. Indeed, it was often impossible to provide anything but a
very loose definition for tissues from different taxa with
the same name (e.g. mammalian and invertebrate trachea
are both involved in the respiratory system, and this is
explicit in the terminology). The definitions that are provided cover tissues that may be unfamiliar (e.g. phyllid)
or whose meaning is slightly technical (e.g. mesonephros
– adult).
We explicitly decided not to use the CARO upper level
anatomy ontology (Haendel et al., 2007) as it is not intuitive to the biologist and is therefore not useful for use in
annotation tools or browsing data, and is actually intended
for use as a template in developing anatomy ontologies
rather than for representing multi-species mappings. We
also decided not to adopt the view that multiple parentage
of terms is undesirable as we are not trying to represent
full anatomical knowledge, rather to produce a resource to
aid data integration pragmatically, and biologists intuitively comprehend multiple parentage as is, for example,
present in the Gene Ontology (Ashburner, et al., 2000).
Validation
Prior to the construction of the MAT terminology, the
ArrayExpress user supplied annotation of ‘OrganismPart’
comprising 817 unique terms used in the annotation of
>60,000 samples obtained from >200 species was
mapped to multiple anatomy ontologies using a Perl implementation of the MetaPhone ‘sound-alike’ algorithm
(Phillips, 1990). The FMA was found to provide the highest coverage of all existing anatomy ontologies, but still
covered only 38% of ArrayExpress anatomical annotations. MAT was mapped to the ArrayExpress annotations
three times during the development of the MAT terminology and the results curated to identify categories of coverage. The final coverage is ~39%.This figure is comparable with the FMA, and uses only 400 terms to achieve
the same coverage. The FMA in contrast contains 25,000
terms and is far less tractable in the context of annotation
tools and usability for the general biomedical scientist.
Format
The MAT terminology uses the OBO (oboedit.org) flat
file format which allows SKOS relationships, and was
constructed using the COBrA editor which has good annotation capabilities that facilitate the mapping of properties such as identifiers and synonyms to MAT terms
(www.xspan.org/cobra).
3
J. Bard et al.
3: DISCUSSION
ACKNOWLEDGEMENTS
The aim of the MAT controlled vocabulary is not only to
produce a standard terminology which can be assigned to
any anatomical parts from any organism, but to provide
primary search terms for those interested in accessing
tissue-associated data. It is intended as a way of integrating data and allowing interoperation between many ontologies.
Thanks to Dawn Field for comments on the manuscript,
and to Dawn, Stuart Aitken, Nick Kruger, Robert Stevens
and Steve Taylor for discussions.
MAT is also intended to help with the strategy of determining the molecular basis of some process in one organism by using information relevant to its development and
function gleaned from other organisms. Here, MAT provides candidate tissues and identifiers, although MAT
tissue groupings may or may not be viewed as equivalent
in any particular context, and the onus is on the user to
choose which tissues may be relevant to their own and, in
turn, which associated data is helpful.
FUNDING
JB thanks the Leverhulme Trust, HP, JM, TFR are funded
in part by EC grants FELICS (contract number 021902),
EMERALD (project number LSHG-CT-2006-037686),
Gen2Phen (contract number 200754) and by EMBL.
REFERENCES
Ashburner, M, Ball, et al. (2000) Gene ontology: tool for the
unification of biology. The Gene Ontology Consortium, Nat
Genet, 25, 25-29.
Bard JBL (2005) Anatomics: the intersection of anatomy and
bioinformatics. J Anat 206: 1-16.
MAT may also be useful in the wider context: as more
data is being generated and funders and journals require
data to be archived, it becomes impossible for database
curators to keep up with the annotation needed for archiving the files, and indeed, harder for the funding agencies
to be able to provide the necessary financial support. A
practical solution to this problem is that people who deposit material in databases annotate their own data in at least
in part. This has always been difficult to achieve formally
for tissue-associated data, and we hope the use of terminologies such as MAT will be helpful here.
Bard J (2007) Anatomy ontologies for model organisms: the
animals and fungi. In: Burger A, Davidson D, Baldock RA editors. Anatomy Ontologies for Bioinformatics. Springer. pp 3-26.
We expect that the majority of users will be interested in a
limited number of taxa, and an editing tool (e.g. COBrA
or OBO-edit) can be used to select only the tissues for
particular organisms. Note that MAT does not seek to
replace the existing ontologies. A more common problem
may be that MAT’s granularity may be too coarse, and
terms may need to be added. This could be solved by using a species-specific ontology, free text, or evolving the
MAT for a specific groups needs. Suggestions, criticisms
and requests should be emailed to [email protected].
Haudry Y, Berube H, et al. (2008). 4DXpress: a database for
cross-species expression pattern comparisons. Nuc Acids Res
36: D847-53.
The MAT terminology does not address the issue of developing an all-encompassing multi-species ontology that
precisely describes orthologous anatomical parts across
evolutionary time. This is a much larger task and has been
attempted in the development of the Bilateria ontology
used in the 4DExpress database of developmental gene
expression data (Haudry et al., 2008). We and others are
participating in discussions to make this a more general
effort. We applaud these efforts and hope that MAT will
be useful in the interim.
4
Burger A, Davidson D, et al. (2003) Formalization of mouse
embryo anatomy, Bioinformatics 19: 1-9.
Burger A, Davidson D, et al. (2007) (editors) Anatomy Ontologies for Bioinformatics. Springer. pp 356.
Haendel, MA., Neuhaus, F, et al (2007). CARO – the common
anatomy reference ontology. In: Burger A, Davidson D, Baldock
RA (editors). Anatomy Ontologies for Bioinformatics. Springer.
pp 327-350.
Kelso J, Visagie J, et al (2003). eVOC: a controlled vocabulary
for unifying gene expression data. Genome Res 6A: 1222-30.
Parkinson H, Kapushesky M, et al. (2006). ArrayExpress -a
public database of microarray experiments and gene expression
profiles. Nucl Acids Res 35: D747-750.
Parkinson H, Aitken S, et al. (2004) The SOFG Anatomy Entry
List (SAEL): An Annotation Tool for Functional Genomics Data.
Comp Funct Gen 5: 521-527.
Phillips L (1990) Hanging on the Metaphone. Comput Lang 7:
39-49
Smith B, Ashburner M, et al. (2007). The OBO Foundry: coodinated evolution of ontologies to support biomedical data integration. Nat Biotechnol 25: 1251
Smith, C, Goldsmith, C-A and Eppig, J (2004) The Mammalian
Phenotype Ontology as a tool for annotating, analyzing and
comparing phenotypic information, Genome Biology, 6.R9
Function, Role, and Disposition in Basic Formal Ontology
Robert Arp* and Barry Smith
National Center for Biomedical Ontology (NCBO) and New York State Center of Excellence in Bioinformatics and Life
Sciences, University at Buffalo, [email protected], [email protected]
ABSTRACT
Numerous research groups are now utilizing Basic Formal
Ontology (BFO) as an upper-level framework to assist in the
organization and integration of biomedical information. This
paper provides elucidation of the three BFO categories of
function, role, and disposition, and considers two proposed
sub-categories of artifactual function and biological function.
The motivation is to help advance the coherent treatment of
functions, roles, and dispositions, to help provide the potential for more detailed classification, and to shed light on
BFO’s general structure and use.
1
INTRODUCTION
Many of the members of the Open Biomedical Ontologies
(OBO) Foundry initiative, including the Gene Ontology, the
Foundational Model of Anatomy, the Protein Ontology, and
the Ontology for Biomedical Investigations (http://www.
obofoundry.org/) are utilizing Basic Formal Ontology
(BFO) to assist in the categorization of entities and relationships in their respective domains of research.
Many individuals and groups involved in organizations
such as BioPAX, Science Commons, Ontology Works, AstraZeneca, and the Computer Task Group utilize BFO as
well.
Fig. 2. The occurrent categories of BFO.
BFO:entity
occurrent
processual entity
process
process boundary
process aggregate
fiat process part
processual context
spatiotemporal region
scattered spatiotemporal region
connected spatiotemporal region
spatiotemporal instant
spatiotemporal interval
temporal region
scattered temporal region
connected temporal region
temporal instant
temporal interval
Fig. 1. The continuant categories of BFO.
BFO:entity
continuant
independent continuant
object
object boundary
object aggregate
fiat object part
site
dependent continuant
generically dependent continuant
specifically dependent continuant
quality
realizable entity
function
role
disposition
spatial region
zero-dimensional region
one-dimensional region
two-dimensional region
three-dimensional region
*
Versions of BFO in OBO, OWL and first-order logic formats are
maintained by Holger Stenzhorn at http://www.ifomis.org/bfo.
Definitions and other content taken from there have been modified to provide additional clarity of exposition.
BFO is an upper-level ontology developed to support integration of data obtained through scientific research. It is
deliberately designed to be very small, in order that is
should be able to represent in consistent fashion the upperlevel categories common to domain ontologies developed by
scientists in different domains and at different levels of granularity. BFO adopts a view of reality as comprising (1)
continuants, entities that continue or persist through time,
such as objects, qualities, and functions, and (2) occurrents,
the events or happenings in which continuants participate.
The subtypes of continuant and occurrent represented in
BFO are presented in Figures 1 and 2 (Grenon and Smith,
2004; Smith and Grenon, 2004; http://www.ifomis.unisaarland.de/bfo/).
To whom correspondence should be addressed.
1
Arp & Smith
2
FUNCTION, ROLE, AND DISPOSITION
Use of the term ‘function’ is common in descriptions of
molecular and cellular processes, as in assertions such as:
• the function of the kidney of Mus musculus is to filter
out waste and water which become urine,
• Arabidopsis thaliana has a multifunctional protein
• there are several folD bifunctional proteins in Campylobacter jejuni.
Functions thus play a central role in the Gene Ontology
(http://www.geneontology.org/).
What, however, of the non-biological functions of artifacts such as screwdrivers, microplates, or pycnometers?
Are there both designed (artifactual) and natural (biological)
functions, representing distinct subtypes of the more general
category of BFO:function?
A related issue is that of the use of the terms ‘function’
and ‘role’. These are distinguished by BFO as representing
two distinct categories (Figure 1), but outside BFO circles
they are often used interchangeably, as when function is
defined as ‘the role that a structure plays in the processes of
a living thing’. Analogous difficulties arise with regard to
the terms ‘disposition’ and ‘tendency’, as in: ‘blood has the
tendency or disposition to coagulate’, ‘a hemophiliac has the
disposition or tendency to bleed an abnormally large amount
of blood’, and ‘that patient has suicidal dispositions or tendencies’.
In this paper, we attempt to elucidate the categories of
function, role, and disposition in BFO. We also describe two
sub-type categories of function, the artifactual and the biological, and provide definitions for each.
Within the context of BFO, one should correctly state:
• the (or a) function of the heart is to pump blood
• the role of the surrogate is to stand in for the patient
• blood has the disposition to coagulate
• that patient has suicidal tendencies
To see why this is so, we need first to consider BFO’s more
general approach to classification.
In BFO, all entities are divided into continuants and occurrents; continuants in turn are divided into independent
and dependent. Independent continuants are things (the objects we see around us every day) in which dependent continuants—such as qualities, functions, roles, dispositions—
can inhere.
Dependent continuants stand to their bearers in the relation of existential dependence: in order for them to exist,
some other (independent) entity must exist. For example,
instances of qualities such as round and red are dependent
continuants in that they cannot exist without being qualities
of some independent continuant such as a ball or a clown’s
nose. So too, functions, roles, and dispositions exist only
insofar as they are functions, roles, and dispositions of some
(one or more) independent continuant. The function of my
2
heart is an instance of the BFO type function, and so also is
the function of your heart.
One major subcategory of dependent continuants in BFO
is that of realizable entity. Realizable entities are defined by
the fact that they can be realized (manifested, actualized,
executed) in occurrents of corresponding sorts. Examples of
realizable entity types include: the function of the liver to
store glycogen, the role of being a doctor, the disposition of
metal to conduct electricity.
Realizable entities are entities of a type whose instances
are typically such that in the course of their existence they
contain periods of actualization, when they are manifested
through processes in which their bearers participate. They
may also exhibit periods of dormancy where they exist by
inhering in their bearers, but are not manifested, as for example, in the case of certain diseases. Some realizables,
such as the function of a sperm to penetrate an ovum, may
be such that they can be manifested only once in their lifetime; or, as again in the case of sperm, they are realized only
in very rare cases.
We are now in a position where we can define function,
role, and disposition.
2.1
Function
A function f is
(1) a realizable dependent continuant.
Thus,
(2) it has a bearer, which is an independent continuant,
and
(3) it is of a type instances of which typically have realizations; each realization is
a. a process in which the bearer is participant
b. that occurs in virtue of the bearer’s physical makeup,
c. and this physical make-up in something which that
bearer possesses because of how it came into being.
Examples include: the function of a birth canal to enable
transport and the function of a hammer to drive in nails. The
process under a. may be specified further as an end-directed
activity, by which we mean in the biological case something
like: an activity that helps to realize the characteristic physiology and life pattern for an organism of the relevant type.
Each function has a bearer with a physical structure which,
in the biological case, the bearer has naturally evolved to
have (as in a hypothalamus secreting hormones) or, in the
artifact case, the bearer has been constructed to have (as in
an Erlenmeyer flask designed to hold liquid) (Ariew and
Perlman, 2002).
It is not accidental or arbitrary that a given eye has the
function to see or that a given screwdriver have been designed and constructed with the function: to fasten screws.
Rather, these functions are integral to these entities in virtue
Function, Role, and Disposition in Basic Formal Ontology
of the fact that the latter have evolved or been constructed to
have a corresponding physical structure.
If a continuant has a function, then it is built to exercise
this function reliably on the basis of this physical structure.
But again: a function is not in every case exercised or manifested. Its bearer may be broken; it may never be in the right
kind of context. Hence, when we say that a given structure
is designed in such a way as to bring about a certain end
reliably, then this reliability presupposes the fulfillment of
certain conditions, for example of an environmental sort.
On the level of instances, this can be stated as: if f is the
function of c, then (in normal circumstances), c exercises f.
On the level of universals, as: if F is the function universal exemplified by instances of the independent continuant
universal C, then (in normal circumstances) instances of C
participate in process instances which are realizations of F.
The implications of this analysis for the treatment of functions in the Gene Ontology are outlined in Hill, Smith,
McAndrews-Hill, and Blake (2008).
2.2
Role
In contrast to function, role is a realizable entity whose manifestation brings about some result or end that is not typical
of its bearer in virtue of the latter’s physical structure. Rather, the role is played by an instance of the corresponding
kind of continuant entity because this entity is in some special natural, social, or institutional set of circumstances
(http://www.ifomis.org/bfo).
Examples include: the role of a chemical compound to
serve as analyte in an experiment, the role of penicillin in
the treatment of a disease, the role of bacteria in causing
infection, the role of a person as student or surgeon.
What is crucial for understanding a role—as distinct from
a function—is that it is a realizable entity that an independent continuant can take on, but that it is not a reflection of
the in-built physical structure of that independent continuant. Certain strains of Escherichia coli bacteria have the
role of pathogen when introduced into the gut of an animal,
but they do not have this role when merely floating around
in a pool of water. A heart has the function of pumping
blood; but in certain circumstances that same heart can play
the role of dinner for the lion.
Roles are optional, and they often involve social ascription. This is why a person can play the role of being a lawyer or a surrogate to a patient, but it is not necessary for
persons that they be lawyers or surrogates.
So, when researchers are considering whether some realizable entity is a function or a role, the question to ask is
this: Is the realizable entity such that its typical manifestations are based upon its physical structure? If so, then it is a
function. Or, is the realizable entity such that its typical manifestation is a reflection of surrounding circumstances, especially those involving social ascription, which are optional? If so, then it is a role.
From this perspective, it is incorrect to make assertions
such as:
• the role of the heart is to pump blood;
• driving nails is a role that this hammer fulfills;
• the function of the surrogate is to stand in for the patient;
• the function of James is to serve as my servant.
2.3
Disposition versus Tendency
It is common to find researchers making claims like: ‘water
has the disposition to rise in a tube’, ‘Carbon-10 has a disposition to decay to Boron-10’, and ‘the cell wall is disposed to filter chemicals in endocitosis and exocitosis.’ A
disposition is a realizable dependent continuant that typically causes a specific process in the object in which it inheres
when the object is introduced into certain specific circumstances. In addition, these processes occur as a result of the
object’s physical structure (Jansen, 2007).
A disposition invariably leads to a certain result given
certain circumstances. Consider: the disposition of a car
windshield to break if struck with a sledgehammer moving
at 100 feet per second; the disposition of a cell to become
diploid following mitosis; the disposition of a magnet to
produce an electrical field.
Contrasted with a disposition is a tendency, which is a
realizable dependent continuant that potentially (not invariably or definitely) causes a specific process in the object in
which it inheres when the object is introduced into certain
specific circumstances as a result of the object’s physical
structure property.
Examples include: the tendency on the part of a hemophiliac to bleed an abnormally large amounts of blood and the
tendency on the part of a person who smokes two packs of
cigarettes a day throughout adulthood to die of a disease at a
below average age. A patient may have a tendency, and not
a disposition, to commit suicide; while a crystal vase has a
disposition, and not a tendency, to break when it hits the
ground after being dropped from a tall building. We are
referring to tendencies when we refer to genetic and other
risk factors for specific diseases.
3
TWO SUB-CATEGORIES OF FUNCTION
It is possible that BFO has failed to recognize categories or
sub-categories of entities existing in reality. The ontology is,
however, developed on the basis of a principle of scientific
fallibilism (Grenon and Smith, 2004). Thus, it is possible
that future research in ontology or in the natural sciences
will reveal the need for an expansion or restructuring of the
categories that BFO recognizes.
In its present form, BFO categories are those included in
the taxonomic hierarchy illustrated in Figures 1 and 2
above. However, we are exploring the possibility of introducing two sub-categories under function, namely artifac3
Arp & Smith
tual function and biological function, as illustrated in Figure
3.
We are also exploring the question of whether to include
tendency as a further sub-category within the ontology.
Fig. 3. Two proposed sub-categories of function in BFO.
BFO:realizable entity
function
artifactual function
biological function
role
disposition
3.1
Artifactual Function
An artifactual function is a function which inheres in an
independent continuant that exists, and has the physical
structure which it has, because it has been designed and
made intentionally (typically by one or more human beings)
to function in a certain way and does indeed reliably function in this way (Lind, 1994; Dipert, 1993).
Examples include: the function of a pycnometer to hold
liquid, the function of a fan to circulate air, and the function
of a Bunsen burner to produce a flame.
3.2
Biological Function
A biological function is a function which inheres in an independent continuant that is (i) part of an organism and (ii)
exists and has the physical structure it has as a result of the
coordinated expression of that organism’s structural genes
(Rosse and Mejino, 2003). The manifestations of a function
of this sort form part of the life of the organism.
Examples include: the function of a mitochondrion in the
production of ATP and the function of the wax-producing
mirror gland of the worker honey bee to produce beeswax.
The manifestations of biological functions are not in
every case beneficial to the survival of the corresponding
organism. (Consider the case of organisms that die when
they reproduce, like Arabis laevigata and Octopus lutens.)
Rather, they are (in typical environments) such as to contribute to the realization by an organism of a life that is typical
or characteristic for an organism of its kind.
It is an open question whether the dichotomy between biological and artifactual function should or should not be
included as an addition to BFO, or reflected rather in the
creation of two new domain ontologies of artifactual and of
biological functions. The latter has already been proposed as
a complement to the GO’s molecular function and biological process ontologies.
4
ACKNOWLEDGEMENTS
We wish to thank Andrew Spear for helpful comments. This
work is funded by the United States National Institutes of
Health (NIH) through the NIH Roadmap for Medical Research, Grant 1 U54 HG004028.
REFERENCES
Allen, C., Bekoff, M., and Lauder, G., eds. (1998) Nature’s Purposes: Analyses of Function and Design in Biology. MIT Press,
Cambridge.
Ariew, A. and Perlman, M. (2002) Introduction. In A. Ariew, R.
Cummins, and M. Perlman, eds., Functions: New Essays in the
Philosophy of Psychology and Biology (pp. 1–7). Oxford University Press, New York.
Arp, R. (2006) Evolution and two popular proposals for the
definition of function. Journal for General Philosophy of
Science, 37, 2–12.
Dipert, R. (1993) Artifacts, Art Works, and Agency. Temple University Press, Philadelphia.
Grenon, P. and Smith, B. (2004) SNAP and SPAN: Towards dynamic spatial ontology. Spatial Cognition and Computation, 4,
99–104.
Hill, D., Smith, B., McAndrews-Hill, M., and Blake, J. (2008)
Gene Ontology annotations: What they mean and where they
come from. BMC Bioinformatics, 9, S2.
Hunter, L. (2009) An Introduction to Molecular Biology for Computer Scientists. MIT Press, Cambridge, MA. In preparation.
Jansen, L. (2007) Tendencies and other realizables in medical information sciences. Available at: http://ontology.buffalo.edu/
bfo/Tendencies.pdf.
Lind, M. (1994) Modeling goals and functions of complex industrial plants. Applied Artificial Intelligence, 8, 259–283.
Perlman, M. (2004) The modern philosophical resurrection of teleology. The Monist, 87, 3–51.
Rosse, C. and Mejino, J. (2003) A reference ontology for bioinformatics: The Foundational Model of Anatomy. Journal of
Biomedical Informatics, 36, 478–500.
Smith, B. and Grenon, P. (2004) The cornucopia of formalontological relations. Dialectica, 58, 279–296.
Smith, B., Kusnierczyk, W., Schober, D., and Ceusters, W. (2006)
Towards a reference terminology for ontology research and
development in the biomedical domain. Proceedings of KRMED, 1, 1–10.
Smith et al. (2007) The OBO Foundry: Coordinated evolution of
ontologies to support biomedical data integration. Nature Biotechnology, 25, 1251–1255.